35#include "mem/request.hh" 36 37template<class Impl> 38LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt, 39 LSQUnit *lsq_ptr) 40 : Event(&mainEventQueue), inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 41{ 42 this->setFlags(Event::AutoDelete); 43} 44 45template<class Impl> 46void 47LSQUnit<Impl>::WritebackEvent::process() 48{ 49 if (!lsqPtr->isSwitchedOut()) { 50 lsqPtr->writeback(inst, pkt); 51 } 52 delete pkt; 53} 54 55template<class Impl> 56const char * 57LSQUnit<Impl>::WritebackEvent::description() 58{ 59 return "Store writeback event"; 60} 61 62template<class Impl> 63void 64LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 65{ 66 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 67 DynInstPtr inst = state->inst; 68 DPRINTF(IEW, "Writeback event [sn:%lli]\n", inst->seqNum); 69 DPRINTF(Activity, "Activity: Writeback event [sn:%lli]\n", inst->seqNum); 70 71 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum); 72 73 if (isSwitchedOut() || inst->isSquashed()) { 74 delete state; 75 delete pkt; 76 return; 77 } else { 78 if (!state->noWB) { 79 writeback(inst, pkt); 80 } 81 82 if (inst->isStore()) { 83 completeStore(state->idx); 84 } 85 } 86 87 delete state; 88 delete pkt; 89} 90 91template <class Impl> 92Tick 93LSQUnit<Impl>::DcachePort::recvAtomic(PacketPtr pkt) 94{ 95 panic("O3CPU model does not work with atomic mode!"); 96 return curTick; 97} 98 99template <class Impl> 100void 101LSQUnit<Impl>::DcachePort::recvFunctional(PacketPtr pkt) 102{ 103 panic("O3CPU doesn't expect recvFunctional callback!"); 104} 105 106template <class Impl> 107void 108LSQUnit<Impl>::DcachePort::recvStatusChange(Status status) 109{ 110 if (status == RangeChange) 111 return; 112 113 panic("O3CPU doesn't expect recvStatusChange callback!"); 114} 115 116template <class Impl> 117bool 118LSQUnit<Impl>::DcachePort::recvTiming(PacketPtr pkt) 119{ 120 lsq->completeDataAccess(pkt); 121 return true; 122} 123 124template <class Impl> 125void 126LSQUnit<Impl>::DcachePort::recvRetry() 127{ 128 lsq->recvRetry(); 129} 130 131template <class Impl> 132LSQUnit<Impl>::LSQUnit() 133 : loads(0), stores(0), storesToWB(0), stalled(false), 134 isStoreBlocked(false), isLoadBlocked(false), 135 loadBlockedHandled(false) 136{ 137} 138 139template<class Impl> 140void 141LSQUnit<Impl>::init(Params *params, unsigned maxLQEntries, 142 unsigned maxSQEntries, unsigned id) 143{ 144 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 145 146 switchedOut = false; 147 148 lsqID = id; 149 150 // Add 1 for the sentinel entry (they are circular queues). 151 LQEntries = maxLQEntries + 1; 152 SQEntries = maxSQEntries + 1; 153 154 loadQueue.resize(LQEntries); 155 storeQueue.resize(SQEntries); 156 157 loadHead = loadTail = 0; 158 159 storeHead = storeWBIdx = storeTail = 0; 160 161 usedPorts = 0; 162 cachePorts = params->cachePorts; 163 164 mem = params->mem; 165 166 memDepViolator = NULL; 167 168 blockedLoadSeqNum = 0; 169} 170 171template<class Impl> 172void 173LSQUnit<Impl>::setCPU(FullCPU *cpu_ptr) 174{ 175 cpu = cpu_ptr; 176 dcachePort = new DcachePort(cpu, this); 177 178 Port *mem_dport = mem->getPort(""); 179 dcachePort->setPeer(mem_dport); 180 mem_dport->setPeer(dcachePort); 181 182 if (cpu->checker) { 183 cpu->checker->setDcachePort(dcachePort); 184 } 185} 186 187template<class Impl> 188std::string 189LSQUnit<Impl>::name() const 190{ 191 if (Impl::MaxThreads == 1) { 192 return iewStage->name() + ".lsq"; 193 } else { 194 return iewStage->name() + ".lsq.thread." + to_string(lsqID); 195 } 196} 197 198template<class Impl> 199void 200LSQUnit<Impl>::clearLQ() 201{ 202 loadQueue.clear(); 203} 204 205template<class Impl> 206void 207LSQUnit<Impl>::clearSQ() 208{ 209 storeQueue.clear(); 210} 211 212template<class Impl> 213void 214LSQUnit<Impl>::switchOut() 215{ 216 switchedOut = true; 217 for (int i = 0; i < loadQueue.size(); ++i) 218 loadQueue[i] = NULL; 219 220 assert(storesToWB == 0); 221} 222 223template<class Impl> 224void 225LSQUnit<Impl>::takeOverFrom() 226{ 227 switchedOut = false; 228 loads = stores = storesToWB = 0; 229 230 loadHead = loadTail = 0; 231 232 storeHead = storeWBIdx = storeTail = 0; 233 234 usedPorts = 0; 235 236 memDepViolator = NULL; 237 238 blockedLoadSeqNum = 0; 239 240 stalled = false; 241 isLoadBlocked = false; 242 loadBlockedHandled = false; 243} 244 245template<class Impl> 246void 247LSQUnit<Impl>::resizeLQ(unsigned size) 248{ 249 unsigned size_plus_sentinel = size + 1; 250 assert(size_plus_sentinel >= LQEntries); 251 252 if (size_plus_sentinel > LQEntries) { 253 while (size_plus_sentinel > loadQueue.size()) { 254 DynInstPtr dummy; 255 loadQueue.push_back(dummy); 256 LQEntries++; 257 } 258 } else { 259 LQEntries = size_plus_sentinel; 260 } 261 262} 263 264template<class Impl> 265void 266LSQUnit<Impl>::resizeSQ(unsigned size) 267{ 268 unsigned size_plus_sentinel = size + 1; 269 if (size_plus_sentinel > SQEntries) { 270 while (size_plus_sentinel > storeQueue.size()) { 271 SQEntry dummy; 272 storeQueue.push_back(dummy); 273 SQEntries++; 274 } 275 } else { 276 SQEntries = size_plus_sentinel; 277 } 278} 279 280template <class Impl> 281void 282LSQUnit<Impl>::insert(DynInstPtr &inst) 283{ 284 assert(inst->isMemRef()); 285 286 assert(inst->isLoad() || inst->isStore()); 287 288 if (inst->isLoad()) { 289 insertLoad(inst); 290 } else { 291 insertStore(inst); 292 } 293 294 inst->setInLSQ(); 295} 296 297template <class Impl> 298void 299LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst) 300{ 301 assert((loadTail + 1) % LQEntries != loadHead); 302 assert(loads < LQEntries); 303 304 DPRINTF(LSQUnit, "Inserting load PC %#x, idx:%i [sn:%lli]\n", 305 load_inst->readPC(), loadTail, load_inst->seqNum); 306 307 load_inst->lqIdx = loadTail; 308 309 if (stores == 0) { 310 load_inst->sqIdx = -1; 311 } else { 312 load_inst->sqIdx = storeTail; 313 } 314 315 loadQueue[loadTail] = load_inst; 316 317 incrLdIdx(loadTail); 318 319 ++loads; 320} 321 322template <class Impl> 323void 324LSQUnit<Impl>::insertStore(DynInstPtr &store_inst) 325{ 326 // Make sure it is not full before inserting an instruction. 327 assert((storeTail + 1) % SQEntries != storeHead); 328 assert(stores < SQEntries); 329 330 DPRINTF(LSQUnit, "Inserting store PC %#x, idx:%i [sn:%lli]\n", 331 store_inst->readPC(), storeTail, store_inst->seqNum); 332 333 store_inst->sqIdx = storeTail; 334 store_inst->lqIdx = loadTail; 335 336 storeQueue[storeTail] = SQEntry(store_inst); 337 338 incrStIdx(storeTail); 339 340 ++stores; 341} 342 343template <class Impl> 344typename Impl::DynInstPtr 345LSQUnit<Impl>::getMemDepViolator() 346{ 347 DynInstPtr temp = memDepViolator; 348 349 memDepViolator = NULL; 350 351 return temp; 352} 353 354template <class Impl> 355unsigned 356LSQUnit<Impl>::numFreeEntries() 357{ 358 unsigned free_lq_entries = LQEntries - loads; 359 unsigned free_sq_entries = SQEntries - stores; 360 361 // Both the LQ and SQ entries have an extra dummy entry to differentiate 362 // empty/full conditions. Subtract 1 from the free entries. 363 if (free_lq_entries < free_sq_entries) { 364 return free_lq_entries - 1; 365 } else { 366 return free_sq_entries - 1; 367 } 368} 369 370template <class Impl> 371int 372LSQUnit<Impl>::numLoadsReady() 373{ 374 int load_idx = loadHead; 375 int retval = 0; 376 377 while (load_idx != loadTail) { 378 assert(loadQueue[load_idx]); 379 380 if (loadQueue[load_idx]->readyToIssue()) { 381 ++retval; 382 } 383 } 384 385 return retval; 386} 387 388template <class Impl> 389Fault 390LSQUnit<Impl>::executeLoad(DynInstPtr &inst) 391{ 392 // Execute a specific load. 393 Fault load_fault = NoFault; 394 395 DPRINTF(LSQUnit, "Executing load PC %#x, [sn:%lli]\n", 396 inst->readPC(),inst->seqNum); 397 398 load_fault = inst->initiateAcc(); 399 400 // If the instruction faulted, then we need to send it along to commit 401 // without the instruction completing. 402 if (load_fault != NoFault) { 403 // Send this instruction to commit, also make sure iew stage 404 // realizes there is activity. 405 iewStage->instToCommit(inst); 406 iewStage->activityThisCycle(); 407 } 408 409 return load_fault; 410} 411 412template <class Impl> 413Fault 414LSQUnit<Impl>::executeStore(DynInstPtr &store_inst) 415{ 416 using namespace TheISA; 417 // Make sure that a store exists. 418 assert(stores != 0); 419 420 int store_idx = store_inst->sqIdx; 421 422 DPRINTF(LSQUnit, "Executing store PC %#x [sn:%lli]\n", 423 store_inst->readPC(), store_inst->seqNum); 424 425 // Check the recently completed loads to see if any match this store's 426 // address. If so, then we have a memory ordering violation. 427 int load_idx = store_inst->lqIdx; 428 429 Fault store_fault = store_inst->initiateAcc(); 430 431 if (storeQueue[store_idx].size == 0) { 432 DPRINTF(LSQUnit,"Fault on Store PC %#x, [sn:%lli],Size = 0\n", 433 store_inst->readPC(),store_inst->seqNum); 434 435 return store_fault; 436 } 437 438 assert(store_fault == NoFault); 439 440 if (store_inst->isStoreConditional()) { 441 // Store conditionals need to set themselves as able to 442 // writeback if we haven't had a fault by here. 443 storeQueue[store_idx].canWB = true; 444 445 ++storesToWB; 446 } 447 448 if (!memDepViolator) { 449 while (load_idx != loadTail) { 450 // Really only need to check loads that have actually executed 451 // It's safe to check all loads because effAddr is set to 452 // InvalAddr when the dyn inst is created. 453 454 // @todo: For now this is extra conservative, detecting a 455 // violation if the addresses match assuming all accesses 456 // are quad word accesses. 457 458 // @todo: Fix this, magic number being used here 459 if ((loadQueue[load_idx]->effAddr >> 8) == 460 (store_inst->effAddr >> 8)) { 461 // A load incorrectly passed this store. Squash and refetch. 462 // For now return a fault to show that it was unsuccessful. 463 memDepViolator = loadQueue[load_idx]; 464 465 return genMachineCheckFault(); 466 } 467 468 incrLdIdx(load_idx); 469 } 470 471 // If we've reached this point, there was no violation. 472 memDepViolator = NULL; 473 } 474 475 return store_fault; 476} 477 478template <class Impl> 479void 480LSQUnit<Impl>::commitLoad() 481{ 482 assert(loadQueue[loadHead]); 483 484 DPRINTF(LSQUnit, "Committing head load instruction, PC %#x\n", 485 loadQueue[loadHead]->readPC()); 486 487 loadQueue[loadHead] = NULL; 488 489 incrLdIdx(loadHead); 490 491 --loads; 492} 493 494template <class Impl> 495void 496LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 497{ 498 assert(loads == 0 || loadQueue[loadHead]); 499 500 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 501 commitLoad(); 502 } 503} 504 505template <class Impl> 506void 507LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 508{ 509 assert(stores == 0 || storeQueue[storeHead].inst); 510 511 int store_idx = storeHead; 512 513 while (store_idx != storeTail) { 514 assert(storeQueue[store_idx].inst); 515 // Mark any stores that are now committed and have not yet 516 // been marked as able to write back. 517 if (!storeQueue[store_idx].canWB) { 518 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 519 break; 520 } 521 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 522 "%#x [sn:%lli]\n", 523 storeQueue[store_idx].inst->readPC(), 524 storeQueue[store_idx].inst->seqNum); 525 526 storeQueue[store_idx].canWB = true; 527 528 ++storesToWB; 529 } 530 531 incrStIdx(store_idx); 532 } 533} 534 535template <class Impl> 536void 537LSQUnit<Impl>::writebackStores() 538{ 539 while (storesToWB > 0 && 540 storeWBIdx != storeTail && 541 storeQueue[storeWBIdx].inst && 542 storeQueue[storeWBIdx].canWB && 543 usedPorts < cachePorts) { 544 545 if (isStoreBlocked) { 546 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 547 " is blocked!\n"); 548 break; 549 } 550 551 // Store didn't write any data so no need to write it back to 552 // memory. 553 if (storeQueue[storeWBIdx].size == 0) { 554 completeStore(storeWBIdx); 555 556 incrStIdx(storeWBIdx); 557 558 continue; 559 } 560 561 ++usedPorts; 562 563 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 564 incrStIdx(storeWBIdx); 565 566 continue; 567 } 568 569 assert(storeQueue[storeWBIdx].req); 570 assert(!storeQueue[storeWBIdx].committed); 571 572 DynInstPtr inst = storeQueue[storeWBIdx].inst; 573 574 Request *req = storeQueue[storeWBIdx].req; 575 storeQueue[storeWBIdx].committed = true; 576 577 assert(!inst->memData); 578 inst->memData = new uint8_t[64]; 579 memcpy(inst->memData, (uint8_t *)&storeQueue[storeWBIdx].data, 580 req->getSize()); 581 582 PacketPtr data_pkt = new Packet(req, Packet::WriteReq, Packet::Broadcast); 583 data_pkt->dataStatic(inst->memData); 584 585 LSQSenderState *state = new LSQSenderState; 586 state->isLoad = false; 587 state->idx = storeWBIdx; 588 state->inst = inst; 589 data_pkt->senderState = state; 590 591 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%#x " 592 "to Addr:%#x, data:%#x [sn:%lli]\n", 593 storeWBIdx, storeQueue[storeWBIdx].inst->readPC(), 594 req->getPaddr(), *(inst->memData), 595 storeQueue[storeWBIdx].inst->seqNum); 596 597 // @todo: Remove this SC hack once the memory system handles it. 598 if (req->getFlags() & LOCKED) { 599 if (req->getFlags() & UNCACHEABLE) { 600 req->setScResult(2); 601 } else { 602 if (cpu->lockFlag) { 603 req->setScResult(1); 604 } else { 605 req->setScResult(0); 606 // Hack: Instantly complete this store. 607 completeDataAccess(data_pkt); 608 incrStIdx(storeWBIdx); 609 continue; 610 } 611 } 612 } else { 613 // Non-store conditionals do not need a writeback. 614 state->noWB = true; 615 } 616 617 if (!dcachePort->sendTiming(data_pkt)) { 618 // Need to handle becoming blocked on a store. 619 isStoreBlocked = true; 620 621 assert(retryPkt == NULL); 622 retryPkt = data_pkt; 623 } else { 624 storePostSend(data_pkt); 625 } 626 } 627 628 // Not sure this should set it to 0. 629 usedPorts = 0; 630 631 assert(stores >= 0 && storesToWB >= 0); 632} 633 634/*template <class Impl> 635void 636LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 637{ 638 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 639 mshrSeqNums.end(), 640 seqNum); 641 642 if (mshr_it != mshrSeqNums.end()) { 643 mshrSeqNums.erase(mshr_it); 644 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 645 } 646}*/ 647 648template <class Impl> 649void 650LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 651{ 652 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 653 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 654 655 int load_idx = loadTail; 656 decrLdIdx(load_idx); 657 658 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 659 DPRINTF(LSQUnit,"Load Instruction PC %#x squashed, " 660 "[sn:%lli]\n", 661 loadQueue[load_idx]->readPC(), 662 loadQueue[load_idx]->seqNum); 663 664 if (isStalled() && load_idx == stallingLoadIdx) { 665 stalled = false; 666 stallingStoreIsn = 0; 667 stallingLoadIdx = 0; 668 } 669 670 // Clear the smart pointer to make sure it is decremented. 671 loadQueue[load_idx]->squashed = true; 672 loadQueue[load_idx] = NULL; 673 --loads; 674 675 // Inefficient! 676 loadTail = load_idx; 677 678 decrLdIdx(load_idx); 679 } 680 681 if (isLoadBlocked) { 682 if (squashed_num < blockedLoadSeqNum) { 683 isLoadBlocked = false; 684 loadBlockedHandled = false; 685 blockedLoadSeqNum = 0; 686 } 687 } 688 689 int store_idx = storeTail; 690 decrStIdx(store_idx); 691 692 while (stores != 0 && 693 storeQueue[store_idx].inst->seqNum > squashed_num) { 694 // Instructions marked as can WB are already committed. 695 if (storeQueue[store_idx].canWB) { 696 break; 697 } 698 699 DPRINTF(LSQUnit,"Store Instruction PC %#x squashed, " 700 "idx:%i [sn:%lli]\n", 701 storeQueue[store_idx].inst->readPC(), 702 store_idx, storeQueue[store_idx].inst->seqNum); 703 704 // I don't think this can happen. It should have been cleared 705 // by the stalling load. 706 if (isStalled() && 707 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 708 panic("Is stalled should have been cleared by stalling load!\n"); 709 stalled = false; 710 stallingStoreIsn = 0; 711 } 712 713 // Clear the smart pointer to make sure it is decremented. 714 storeQueue[store_idx].inst->squashed = true; 715 storeQueue[store_idx].inst = NULL; 716 storeQueue[store_idx].canWB = 0; 717 718 storeQueue[store_idx].req = NULL; 719 --stores; 720 721 // Inefficient! 722 storeTail = store_idx; 723 724 decrStIdx(store_idx); 725 } 726} 727 728template <class Impl> 729void 730LSQUnit<Impl>::storePostSend(Packet *pkt) 731{ 732 if (isStalled() && 733 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 734 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 735 "load idx:%i\n", 736 stallingStoreIsn, stallingLoadIdx); 737 stalled = false; 738 stallingStoreIsn = 0; 739 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 740 } 741 742 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 743 // The store is basically completed at this time. This 744 // only works so long as the checker doesn't try to 745 // verify the value in memory for stores. 746 storeQueue[storeWBIdx].inst->setCompleted(); 747 if (cpu->checker) { 748 cpu->checker->tick(storeQueue[storeWBIdx].inst); 749 } 750 } 751 752 if (pkt->result != Packet::Success) { 753 DPRINTF(LSQUnit,"D-Cache Write Miss on idx:%i!\n", 754 storeWBIdx); 755 756 DPRINTF(Activity, "Active st accessing mem miss [sn:%lli]\n", 757 storeQueue[storeWBIdx].inst->seqNum); 758 759 //mshrSeqNums.push_back(storeQueue[storeWBIdx].inst->seqNum); 760 761 //DPRINTF(LSQUnit, "Added MSHR. count = %i\n",mshrSeqNums.size()); 762 763 // @todo: Increment stat here. 764 } else { 765 DPRINTF(LSQUnit,"D-Cache: Write Hit on idx:%i !\n", 766 storeWBIdx); 767 768 DPRINTF(Activity, "Active st accessing mem hit [sn:%lli]\n", 769 storeQueue[storeWBIdx].inst->seqNum); 770 } 771 772 incrStIdx(storeWBIdx); 773} 774 775template <class Impl> 776void 777LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt) 778{ 779 iewStage->wakeCPU(); 780 781 // Squashed instructions do not need to complete their access. 782 if (inst->isSquashed()) { 783 assert(!inst->isStore()); 784 return; 785 } 786 787 if (!inst->isExecuted()) { 788 inst->setExecuted(); 789 790 // Complete access to copy data to proper place. 791 inst->completeAcc(pkt); 792 } 793 794 // Need to insert instruction into queue to commit 795 iewStage->instToCommit(inst); 796 797 iewStage->activityThisCycle(); 798} 799 800template <class Impl> 801void 802LSQUnit<Impl>::completeStore(int store_idx) 803{ 804 assert(storeQueue[store_idx].inst); 805 storeQueue[store_idx].completed = true; 806 --storesToWB; 807 // A bit conservative because a store completion may not free up entries, 808 // but hopefully avoids two store completions in one cycle from making 809 // the CPU tick twice. 810 cpu->activityThisCycle(); 811 812 if (store_idx == storeHead) { 813 do { 814 incrStIdx(storeHead); 815 816 --stores; 817 } while (storeQueue[storeHead].completed && 818 storeHead != storeTail); 819 820 iewStage->updateLSQNextCycle = true; 821 } 822 823 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 824 "idx:%i\n", 825 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 826 827 if (isStalled() && 828 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 829 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 830 "load idx:%i\n", 831 stallingStoreIsn, stallingLoadIdx); 832 stalled = false; 833 stallingStoreIsn = 0; 834 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 835 } 836 837 storeQueue[store_idx].inst->setCompleted(); 838 839 // Tell the checker we've completed this instruction. Some stores 840 // may get reported twice to the checker, but the checker can 841 // handle that case. 842 if (cpu->checker) { 843 cpu->checker->tick(storeQueue[store_idx].inst); 844 } 845} 846 847template <class Impl> 848void 849LSQUnit<Impl>::recvRetry() 850{ 851 if (isStoreBlocked) { 852 assert(retryPkt != NULL); 853 854 if (dcachePort->sendTiming(retryPkt)) { 855 storePostSend(retryPkt); 856 retryPkt = NULL; 857 isStoreBlocked = false; 858 } else { 859 // Still blocked! 860 } 861 } else if (isLoadBlocked) { 862 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, " 863 "no need to resend packet.\n"); 864 } else { 865 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n"); 866 } 867} 868 869template <class Impl> 870inline void 871LSQUnit<Impl>::incrStIdx(int &store_idx) 872{ 873 if (++store_idx >= SQEntries) 874 store_idx = 0; 875} 876 877template <class Impl> 878inline void 879LSQUnit<Impl>::decrStIdx(int &store_idx) 880{ 881 if (--store_idx < 0) 882 store_idx += SQEntries; 883} 884 885template <class Impl> 886inline void 887LSQUnit<Impl>::incrLdIdx(int &load_idx) 888{ 889 if (++load_idx >= LQEntries) 890 load_idx = 0; 891} 892 893template <class Impl> 894inline void 895LSQUnit<Impl>::decrLdIdx(int &load_idx) 896{ 897 if (--load_idx < 0) 898 load_idx += LQEntries; 899} 900 901template <class Impl> 902void 903LSQUnit<Impl>::dumpInsts() 904{ 905 cprintf("Load store queue: Dumping instructions.\n"); 906 cprintf("Load queue size: %i\n", loads); 907 cprintf("Load queue: "); 908 909 int load_idx = loadHead; 910 911 while (load_idx != loadTail && loadQueue[load_idx]) { 912 cprintf("%#x ", loadQueue[load_idx]->readPC()); 913 914 incrLdIdx(load_idx); 915 } 916 917 cprintf("Store queue size: %i\n", stores); 918 cprintf("Store queue: "); 919 920 int store_idx = storeHead; 921 922 while (store_idx != storeTail && storeQueue[store_idx].inst) { 923 cprintf("%#x ", storeQueue[store_idx].inst->readPC()); 924 925 incrStIdx(store_idx); 926 } 927 928 cprintf("\n"); 929}
| 36#include "mem/request.hh" 37 38template<class Impl> 39LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt, 40 LSQUnit *lsq_ptr) 41 : Event(&mainEventQueue), inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 42{ 43 this->setFlags(Event::AutoDelete); 44} 45 46template<class Impl> 47void 48LSQUnit<Impl>::WritebackEvent::process() 49{ 50 if (!lsqPtr->isSwitchedOut()) { 51 lsqPtr->writeback(inst, pkt); 52 } 53 delete pkt; 54} 55 56template<class Impl> 57const char * 58LSQUnit<Impl>::WritebackEvent::description() 59{ 60 return "Store writeback event"; 61} 62 63template<class Impl> 64void 65LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 66{ 67 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 68 DynInstPtr inst = state->inst; 69 DPRINTF(IEW, "Writeback event [sn:%lli]\n", inst->seqNum); 70 DPRINTF(Activity, "Activity: Writeback event [sn:%lli]\n", inst->seqNum); 71 72 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum); 73 74 if (isSwitchedOut() || inst->isSquashed()) { 75 delete state; 76 delete pkt; 77 return; 78 } else { 79 if (!state->noWB) { 80 writeback(inst, pkt); 81 } 82 83 if (inst->isStore()) { 84 completeStore(state->idx); 85 } 86 } 87 88 delete state; 89 delete pkt; 90} 91 92template <class Impl> 93Tick 94LSQUnit<Impl>::DcachePort::recvAtomic(PacketPtr pkt) 95{ 96 panic("O3CPU model does not work with atomic mode!"); 97 return curTick; 98} 99 100template <class Impl> 101void 102LSQUnit<Impl>::DcachePort::recvFunctional(PacketPtr pkt) 103{ 104 panic("O3CPU doesn't expect recvFunctional callback!"); 105} 106 107template <class Impl> 108void 109LSQUnit<Impl>::DcachePort::recvStatusChange(Status status) 110{ 111 if (status == RangeChange) 112 return; 113 114 panic("O3CPU doesn't expect recvStatusChange callback!"); 115} 116 117template <class Impl> 118bool 119LSQUnit<Impl>::DcachePort::recvTiming(PacketPtr pkt) 120{ 121 lsq->completeDataAccess(pkt); 122 return true; 123} 124 125template <class Impl> 126void 127LSQUnit<Impl>::DcachePort::recvRetry() 128{ 129 lsq->recvRetry(); 130} 131 132template <class Impl> 133LSQUnit<Impl>::LSQUnit() 134 : loads(0), stores(0), storesToWB(0), stalled(false), 135 isStoreBlocked(false), isLoadBlocked(false), 136 loadBlockedHandled(false) 137{ 138} 139 140template<class Impl> 141void 142LSQUnit<Impl>::init(Params *params, unsigned maxLQEntries, 143 unsigned maxSQEntries, unsigned id) 144{ 145 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 146 147 switchedOut = false; 148 149 lsqID = id; 150 151 // Add 1 for the sentinel entry (they are circular queues). 152 LQEntries = maxLQEntries + 1; 153 SQEntries = maxSQEntries + 1; 154 155 loadQueue.resize(LQEntries); 156 storeQueue.resize(SQEntries); 157 158 loadHead = loadTail = 0; 159 160 storeHead = storeWBIdx = storeTail = 0; 161 162 usedPorts = 0; 163 cachePorts = params->cachePorts; 164 165 mem = params->mem; 166 167 memDepViolator = NULL; 168 169 blockedLoadSeqNum = 0; 170} 171 172template<class Impl> 173void 174LSQUnit<Impl>::setCPU(FullCPU *cpu_ptr) 175{ 176 cpu = cpu_ptr; 177 dcachePort = new DcachePort(cpu, this); 178 179 Port *mem_dport = mem->getPort(""); 180 dcachePort->setPeer(mem_dport); 181 mem_dport->setPeer(dcachePort); 182 183 if (cpu->checker) { 184 cpu->checker->setDcachePort(dcachePort); 185 } 186} 187 188template<class Impl> 189std::string 190LSQUnit<Impl>::name() const 191{ 192 if (Impl::MaxThreads == 1) { 193 return iewStage->name() + ".lsq"; 194 } else { 195 return iewStage->name() + ".lsq.thread." + to_string(lsqID); 196 } 197} 198 199template<class Impl> 200void 201LSQUnit<Impl>::clearLQ() 202{ 203 loadQueue.clear(); 204} 205 206template<class Impl> 207void 208LSQUnit<Impl>::clearSQ() 209{ 210 storeQueue.clear(); 211} 212 213template<class Impl> 214void 215LSQUnit<Impl>::switchOut() 216{ 217 switchedOut = true; 218 for (int i = 0; i < loadQueue.size(); ++i) 219 loadQueue[i] = NULL; 220 221 assert(storesToWB == 0); 222} 223 224template<class Impl> 225void 226LSQUnit<Impl>::takeOverFrom() 227{ 228 switchedOut = false; 229 loads = stores = storesToWB = 0; 230 231 loadHead = loadTail = 0; 232 233 storeHead = storeWBIdx = storeTail = 0; 234 235 usedPorts = 0; 236 237 memDepViolator = NULL; 238 239 blockedLoadSeqNum = 0; 240 241 stalled = false; 242 isLoadBlocked = false; 243 loadBlockedHandled = false; 244} 245 246template<class Impl> 247void 248LSQUnit<Impl>::resizeLQ(unsigned size) 249{ 250 unsigned size_plus_sentinel = size + 1; 251 assert(size_plus_sentinel >= LQEntries); 252 253 if (size_plus_sentinel > LQEntries) { 254 while (size_plus_sentinel > loadQueue.size()) { 255 DynInstPtr dummy; 256 loadQueue.push_back(dummy); 257 LQEntries++; 258 } 259 } else { 260 LQEntries = size_plus_sentinel; 261 } 262 263} 264 265template<class Impl> 266void 267LSQUnit<Impl>::resizeSQ(unsigned size) 268{ 269 unsigned size_plus_sentinel = size + 1; 270 if (size_plus_sentinel > SQEntries) { 271 while (size_plus_sentinel > storeQueue.size()) { 272 SQEntry dummy; 273 storeQueue.push_back(dummy); 274 SQEntries++; 275 } 276 } else { 277 SQEntries = size_plus_sentinel; 278 } 279} 280 281template <class Impl> 282void 283LSQUnit<Impl>::insert(DynInstPtr &inst) 284{ 285 assert(inst->isMemRef()); 286 287 assert(inst->isLoad() || inst->isStore()); 288 289 if (inst->isLoad()) { 290 insertLoad(inst); 291 } else { 292 insertStore(inst); 293 } 294 295 inst->setInLSQ(); 296} 297 298template <class Impl> 299void 300LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst) 301{ 302 assert((loadTail + 1) % LQEntries != loadHead); 303 assert(loads < LQEntries); 304 305 DPRINTF(LSQUnit, "Inserting load PC %#x, idx:%i [sn:%lli]\n", 306 load_inst->readPC(), loadTail, load_inst->seqNum); 307 308 load_inst->lqIdx = loadTail; 309 310 if (stores == 0) { 311 load_inst->sqIdx = -1; 312 } else { 313 load_inst->sqIdx = storeTail; 314 } 315 316 loadQueue[loadTail] = load_inst; 317 318 incrLdIdx(loadTail); 319 320 ++loads; 321} 322 323template <class Impl> 324void 325LSQUnit<Impl>::insertStore(DynInstPtr &store_inst) 326{ 327 // Make sure it is not full before inserting an instruction. 328 assert((storeTail + 1) % SQEntries != storeHead); 329 assert(stores < SQEntries); 330 331 DPRINTF(LSQUnit, "Inserting store PC %#x, idx:%i [sn:%lli]\n", 332 store_inst->readPC(), storeTail, store_inst->seqNum); 333 334 store_inst->sqIdx = storeTail; 335 store_inst->lqIdx = loadTail; 336 337 storeQueue[storeTail] = SQEntry(store_inst); 338 339 incrStIdx(storeTail); 340 341 ++stores; 342} 343 344template <class Impl> 345typename Impl::DynInstPtr 346LSQUnit<Impl>::getMemDepViolator() 347{ 348 DynInstPtr temp = memDepViolator; 349 350 memDepViolator = NULL; 351 352 return temp; 353} 354 355template <class Impl> 356unsigned 357LSQUnit<Impl>::numFreeEntries() 358{ 359 unsigned free_lq_entries = LQEntries - loads; 360 unsigned free_sq_entries = SQEntries - stores; 361 362 // Both the LQ and SQ entries have an extra dummy entry to differentiate 363 // empty/full conditions. Subtract 1 from the free entries. 364 if (free_lq_entries < free_sq_entries) { 365 return free_lq_entries - 1; 366 } else { 367 return free_sq_entries - 1; 368 } 369} 370 371template <class Impl> 372int 373LSQUnit<Impl>::numLoadsReady() 374{ 375 int load_idx = loadHead; 376 int retval = 0; 377 378 while (load_idx != loadTail) { 379 assert(loadQueue[load_idx]); 380 381 if (loadQueue[load_idx]->readyToIssue()) { 382 ++retval; 383 } 384 } 385 386 return retval; 387} 388 389template <class Impl> 390Fault 391LSQUnit<Impl>::executeLoad(DynInstPtr &inst) 392{ 393 // Execute a specific load. 394 Fault load_fault = NoFault; 395 396 DPRINTF(LSQUnit, "Executing load PC %#x, [sn:%lli]\n", 397 inst->readPC(),inst->seqNum); 398 399 load_fault = inst->initiateAcc(); 400 401 // If the instruction faulted, then we need to send it along to commit 402 // without the instruction completing. 403 if (load_fault != NoFault) { 404 // Send this instruction to commit, also make sure iew stage 405 // realizes there is activity. 406 iewStage->instToCommit(inst); 407 iewStage->activityThisCycle(); 408 } 409 410 return load_fault; 411} 412 413template <class Impl> 414Fault 415LSQUnit<Impl>::executeStore(DynInstPtr &store_inst) 416{ 417 using namespace TheISA; 418 // Make sure that a store exists. 419 assert(stores != 0); 420 421 int store_idx = store_inst->sqIdx; 422 423 DPRINTF(LSQUnit, "Executing store PC %#x [sn:%lli]\n", 424 store_inst->readPC(), store_inst->seqNum); 425 426 // Check the recently completed loads to see if any match this store's 427 // address. If so, then we have a memory ordering violation. 428 int load_idx = store_inst->lqIdx; 429 430 Fault store_fault = store_inst->initiateAcc(); 431 432 if (storeQueue[store_idx].size == 0) { 433 DPRINTF(LSQUnit,"Fault on Store PC %#x, [sn:%lli],Size = 0\n", 434 store_inst->readPC(),store_inst->seqNum); 435 436 return store_fault; 437 } 438 439 assert(store_fault == NoFault); 440 441 if (store_inst->isStoreConditional()) { 442 // Store conditionals need to set themselves as able to 443 // writeback if we haven't had a fault by here. 444 storeQueue[store_idx].canWB = true; 445 446 ++storesToWB; 447 } 448 449 if (!memDepViolator) { 450 while (load_idx != loadTail) { 451 // Really only need to check loads that have actually executed 452 // It's safe to check all loads because effAddr is set to 453 // InvalAddr when the dyn inst is created. 454 455 // @todo: For now this is extra conservative, detecting a 456 // violation if the addresses match assuming all accesses 457 // are quad word accesses. 458 459 // @todo: Fix this, magic number being used here 460 if ((loadQueue[load_idx]->effAddr >> 8) == 461 (store_inst->effAddr >> 8)) { 462 // A load incorrectly passed this store. Squash and refetch. 463 // For now return a fault to show that it was unsuccessful. 464 memDepViolator = loadQueue[load_idx]; 465 466 return genMachineCheckFault(); 467 } 468 469 incrLdIdx(load_idx); 470 } 471 472 // If we've reached this point, there was no violation. 473 memDepViolator = NULL; 474 } 475 476 return store_fault; 477} 478 479template <class Impl> 480void 481LSQUnit<Impl>::commitLoad() 482{ 483 assert(loadQueue[loadHead]); 484 485 DPRINTF(LSQUnit, "Committing head load instruction, PC %#x\n", 486 loadQueue[loadHead]->readPC()); 487 488 loadQueue[loadHead] = NULL; 489 490 incrLdIdx(loadHead); 491 492 --loads; 493} 494 495template <class Impl> 496void 497LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 498{ 499 assert(loads == 0 || loadQueue[loadHead]); 500 501 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 502 commitLoad(); 503 } 504} 505 506template <class Impl> 507void 508LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 509{ 510 assert(stores == 0 || storeQueue[storeHead].inst); 511 512 int store_idx = storeHead; 513 514 while (store_idx != storeTail) { 515 assert(storeQueue[store_idx].inst); 516 // Mark any stores that are now committed and have not yet 517 // been marked as able to write back. 518 if (!storeQueue[store_idx].canWB) { 519 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 520 break; 521 } 522 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 523 "%#x [sn:%lli]\n", 524 storeQueue[store_idx].inst->readPC(), 525 storeQueue[store_idx].inst->seqNum); 526 527 storeQueue[store_idx].canWB = true; 528 529 ++storesToWB; 530 } 531 532 incrStIdx(store_idx); 533 } 534} 535 536template <class Impl> 537void 538LSQUnit<Impl>::writebackStores() 539{ 540 while (storesToWB > 0 && 541 storeWBIdx != storeTail && 542 storeQueue[storeWBIdx].inst && 543 storeQueue[storeWBIdx].canWB && 544 usedPorts < cachePorts) { 545 546 if (isStoreBlocked) { 547 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 548 " is blocked!\n"); 549 break; 550 } 551 552 // Store didn't write any data so no need to write it back to 553 // memory. 554 if (storeQueue[storeWBIdx].size == 0) { 555 completeStore(storeWBIdx); 556 557 incrStIdx(storeWBIdx); 558 559 continue; 560 } 561 562 ++usedPorts; 563 564 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 565 incrStIdx(storeWBIdx); 566 567 continue; 568 } 569 570 assert(storeQueue[storeWBIdx].req); 571 assert(!storeQueue[storeWBIdx].committed); 572 573 DynInstPtr inst = storeQueue[storeWBIdx].inst; 574 575 Request *req = storeQueue[storeWBIdx].req; 576 storeQueue[storeWBIdx].committed = true; 577 578 assert(!inst->memData); 579 inst->memData = new uint8_t[64]; 580 memcpy(inst->memData, (uint8_t *)&storeQueue[storeWBIdx].data, 581 req->getSize()); 582 583 PacketPtr data_pkt = new Packet(req, Packet::WriteReq, Packet::Broadcast); 584 data_pkt->dataStatic(inst->memData); 585 586 LSQSenderState *state = new LSQSenderState; 587 state->isLoad = false; 588 state->idx = storeWBIdx; 589 state->inst = inst; 590 data_pkt->senderState = state; 591 592 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%#x " 593 "to Addr:%#x, data:%#x [sn:%lli]\n", 594 storeWBIdx, storeQueue[storeWBIdx].inst->readPC(), 595 req->getPaddr(), *(inst->memData), 596 storeQueue[storeWBIdx].inst->seqNum); 597 598 // @todo: Remove this SC hack once the memory system handles it. 599 if (req->getFlags() & LOCKED) { 600 if (req->getFlags() & UNCACHEABLE) { 601 req->setScResult(2); 602 } else { 603 if (cpu->lockFlag) { 604 req->setScResult(1); 605 } else { 606 req->setScResult(0); 607 // Hack: Instantly complete this store. 608 completeDataAccess(data_pkt); 609 incrStIdx(storeWBIdx); 610 continue; 611 } 612 } 613 } else { 614 // Non-store conditionals do not need a writeback. 615 state->noWB = true; 616 } 617 618 if (!dcachePort->sendTiming(data_pkt)) { 619 // Need to handle becoming blocked on a store. 620 isStoreBlocked = true; 621 622 assert(retryPkt == NULL); 623 retryPkt = data_pkt; 624 } else { 625 storePostSend(data_pkt); 626 } 627 } 628 629 // Not sure this should set it to 0. 630 usedPorts = 0; 631 632 assert(stores >= 0 && storesToWB >= 0); 633} 634 635/*template <class Impl> 636void 637LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 638{ 639 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 640 mshrSeqNums.end(), 641 seqNum); 642 643 if (mshr_it != mshrSeqNums.end()) { 644 mshrSeqNums.erase(mshr_it); 645 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 646 } 647}*/ 648 649template <class Impl> 650void 651LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 652{ 653 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 654 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 655 656 int load_idx = loadTail; 657 decrLdIdx(load_idx); 658 659 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 660 DPRINTF(LSQUnit,"Load Instruction PC %#x squashed, " 661 "[sn:%lli]\n", 662 loadQueue[load_idx]->readPC(), 663 loadQueue[load_idx]->seqNum); 664 665 if (isStalled() && load_idx == stallingLoadIdx) { 666 stalled = false; 667 stallingStoreIsn = 0; 668 stallingLoadIdx = 0; 669 } 670 671 // Clear the smart pointer to make sure it is decremented. 672 loadQueue[load_idx]->squashed = true; 673 loadQueue[load_idx] = NULL; 674 --loads; 675 676 // Inefficient! 677 loadTail = load_idx; 678 679 decrLdIdx(load_idx); 680 } 681 682 if (isLoadBlocked) { 683 if (squashed_num < blockedLoadSeqNum) { 684 isLoadBlocked = false; 685 loadBlockedHandled = false; 686 blockedLoadSeqNum = 0; 687 } 688 } 689 690 int store_idx = storeTail; 691 decrStIdx(store_idx); 692 693 while (stores != 0 && 694 storeQueue[store_idx].inst->seqNum > squashed_num) { 695 // Instructions marked as can WB are already committed. 696 if (storeQueue[store_idx].canWB) { 697 break; 698 } 699 700 DPRINTF(LSQUnit,"Store Instruction PC %#x squashed, " 701 "idx:%i [sn:%lli]\n", 702 storeQueue[store_idx].inst->readPC(), 703 store_idx, storeQueue[store_idx].inst->seqNum); 704 705 // I don't think this can happen. It should have been cleared 706 // by the stalling load. 707 if (isStalled() && 708 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 709 panic("Is stalled should have been cleared by stalling load!\n"); 710 stalled = false; 711 stallingStoreIsn = 0; 712 } 713 714 // Clear the smart pointer to make sure it is decremented. 715 storeQueue[store_idx].inst->squashed = true; 716 storeQueue[store_idx].inst = NULL; 717 storeQueue[store_idx].canWB = 0; 718 719 storeQueue[store_idx].req = NULL; 720 --stores; 721 722 // Inefficient! 723 storeTail = store_idx; 724 725 decrStIdx(store_idx); 726 } 727} 728 729template <class Impl> 730void 731LSQUnit<Impl>::storePostSend(Packet *pkt) 732{ 733 if (isStalled() && 734 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 735 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 736 "load idx:%i\n", 737 stallingStoreIsn, stallingLoadIdx); 738 stalled = false; 739 stallingStoreIsn = 0; 740 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 741 } 742 743 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 744 // The store is basically completed at this time. This 745 // only works so long as the checker doesn't try to 746 // verify the value in memory for stores. 747 storeQueue[storeWBIdx].inst->setCompleted(); 748 if (cpu->checker) { 749 cpu->checker->tick(storeQueue[storeWBIdx].inst); 750 } 751 } 752 753 if (pkt->result != Packet::Success) { 754 DPRINTF(LSQUnit,"D-Cache Write Miss on idx:%i!\n", 755 storeWBIdx); 756 757 DPRINTF(Activity, "Active st accessing mem miss [sn:%lli]\n", 758 storeQueue[storeWBIdx].inst->seqNum); 759 760 //mshrSeqNums.push_back(storeQueue[storeWBIdx].inst->seqNum); 761 762 //DPRINTF(LSQUnit, "Added MSHR. count = %i\n",mshrSeqNums.size()); 763 764 // @todo: Increment stat here. 765 } else { 766 DPRINTF(LSQUnit,"D-Cache: Write Hit on idx:%i !\n", 767 storeWBIdx); 768 769 DPRINTF(Activity, "Active st accessing mem hit [sn:%lli]\n", 770 storeQueue[storeWBIdx].inst->seqNum); 771 } 772 773 incrStIdx(storeWBIdx); 774} 775 776template <class Impl> 777void 778LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt) 779{ 780 iewStage->wakeCPU(); 781 782 // Squashed instructions do not need to complete their access. 783 if (inst->isSquashed()) { 784 assert(!inst->isStore()); 785 return; 786 } 787 788 if (!inst->isExecuted()) { 789 inst->setExecuted(); 790 791 // Complete access to copy data to proper place. 792 inst->completeAcc(pkt); 793 } 794 795 // Need to insert instruction into queue to commit 796 iewStage->instToCommit(inst); 797 798 iewStage->activityThisCycle(); 799} 800 801template <class Impl> 802void 803LSQUnit<Impl>::completeStore(int store_idx) 804{ 805 assert(storeQueue[store_idx].inst); 806 storeQueue[store_idx].completed = true; 807 --storesToWB; 808 // A bit conservative because a store completion may not free up entries, 809 // but hopefully avoids two store completions in one cycle from making 810 // the CPU tick twice. 811 cpu->activityThisCycle(); 812 813 if (store_idx == storeHead) { 814 do { 815 incrStIdx(storeHead); 816 817 --stores; 818 } while (storeQueue[storeHead].completed && 819 storeHead != storeTail); 820 821 iewStage->updateLSQNextCycle = true; 822 } 823 824 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 825 "idx:%i\n", 826 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 827 828 if (isStalled() && 829 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 830 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 831 "load idx:%i\n", 832 stallingStoreIsn, stallingLoadIdx); 833 stalled = false; 834 stallingStoreIsn = 0; 835 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 836 } 837 838 storeQueue[store_idx].inst->setCompleted(); 839 840 // Tell the checker we've completed this instruction. Some stores 841 // may get reported twice to the checker, but the checker can 842 // handle that case. 843 if (cpu->checker) { 844 cpu->checker->tick(storeQueue[store_idx].inst); 845 } 846} 847 848template <class Impl> 849void 850LSQUnit<Impl>::recvRetry() 851{ 852 if (isStoreBlocked) { 853 assert(retryPkt != NULL); 854 855 if (dcachePort->sendTiming(retryPkt)) { 856 storePostSend(retryPkt); 857 retryPkt = NULL; 858 isStoreBlocked = false; 859 } else { 860 // Still blocked! 861 } 862 } else if (isLoadBlocked) { 863 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, " 864 "no need to resend packet.\n"); 865 } else { 866 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n"); 867 } 868} 869 870template <class Impl> 871inline void 872LSQUnit<Impl>::incrStIdx(int &store_idx) 873{ 874 if (++store_idx >= SQEntries) 875 store_idx = 0; 876} 877 878template <class Impl> 879inline void 880LSQUnit<Impl>::decrStIdx(int &store_idx) 881{ 882 if (--store_idx < 0) 883 store_idx += SQEntries; 884} 885 886template <class Impl> 887inline void 888LSQUnit<Impl>::incrLdIdx(int &load_idx) 889{ 890 if (++load_idx >= LQEntries) 891 load_idx = 0; 892} 893 894template <class Impl> 895inline void 896LSQUnit<Impl>::decrLdIdx(int &load_idx) 897{ 898 if (--load_idx < 0) 899 load_idx += LQEntries; 900} 901 902template <class Impl> 903void 904LSQUnit<Impl>::dumpInsts() 905{ 906 cprintf("Load store queue: Dumping instructions.\n"); 907 cprintf("Load queue size: %i\n", loads); 908 cprintf("Load queue: "); 909 910 int load_idx = loadHead; 911 912 while (load_idx != loadTail && loadQueue[load_idx]) { 913 cprintf("%#x ", loadQueue[load_idx]->readPC()); 914 915 incrLdIdx(load_idx); 916 } 917 918 cprintf("Store queue size: %i\n", stores); 919 cprintf("Store queue: "); 920 921 int store_idx = storeHead; 922 923 while (store_idx != storeTail && storeQueue[store_idx].inst) { 924 cprintf("%#x ", storeQueue[store_idx].inst->readPC()); 925 926 incrStIdx(store_idx); 927 } 928 929 cprintf("\n"); 930}
|