1/* 2 * Copyright (c) 2010 ARM Limited 3 * All rights reserved 4 * 5 * The license below extends only to copyright in the software and shall 6 * not be construed as granting a license to any other intellectual 7 * property including but not limited to intellectual property relating 8 * to a hardware implementation of the functionality of the software 9 * licensed hereunder. You may use the software subject to the license 10 * terms below provided that you ensure that this notice is replicated 11 * unmodified and in its entirety in all distributions of the software, 12 * modified or unmodified, in source code or in binary form. 13 * 14 * Copyright (c) 2004-2005 The Regents of The University of Michigan 15 * All rights reserved. 16 * 17 * Redistribution and use in source and binary forms, with or without 18 * modification, are permitted provided that the following conditions are 19 * met: redistributions of source code must retain the above copyright 20 * notice, this list of conditions and the following disclaimer; 21 * redistributions in binary form must reproduce the above copyright 22 * notice, this list of conditions and the following disclaimer in the 23 * documentation and/or other materials provided with the distribution; 24 * neither the name of the copyright holders nor the names of its 25 * contributors may be used to endorse or promote products derived from 26 * this software without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Authors: Kevin Lim 41 * Korey Sewell 42 */ 43 44#include "arch/locked_mem.hh" 45#include "base/str.hh" 46#include "config/the_isa.hh" 47#include "config/use_checker.hh" 48#include "cpu/o3/lsq.hh" 49#include "cpu/o3/lsq_unit.hh" 50#include "debug/Activity.hh" 51#include "debug/IEW.hh" 52#include "debug/LSQUnit.hh" 53#include "mem/packet.hh" 54#include "mem/request.hh" 55 56#if USE_CHECKER 57#include "cpu/checker/cpu.hh" 58#endif 59 60template<class Impl> 61LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt, 62 LSQUnit *lsq_ptr) 63 : inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 64{ 65 this->setFlags(Event::AutoDelete); 66} 67 68template<class Impl> 69void 70LSQUnit<Impl>::WritebackEvent::process() 71{ 72 if (!lsqPtr->isSwitchedOut()) { 73 lsqPtr->writeback(inst, pkt); 74 } 75 76 if (pkt->senderState) 77 delete pkt->senderState; 78 79 delete pkt->req; 80 delete pkt; 81} 82 83template<class Impl> 84const char * 85LSQUnit<Impl>::WritebackEvent::description() const 86{ 87 return "Store writeback"; 88} 89 90template<class Impl> 91void 92LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 93{ 94 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 95 DynInstPtr inst = state->inst; 96 DPRINTF(IEW, "Writeback event [sn:%lli].\n", inst->seqNum); 97 DPRINTF(Activity, "Activity: Writeback event [sn:%lli].\n", inst->seqNum); 98 99 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum); 100 101 assert(!pkt->wasNacked()); 102 103 // If this is a split access, wait until all packets are received. 104 if (TheISA::HasUnalignedMemAcc && !state->complete()) { 105 delete pkt->req; 106 delete pkt; 107 return; 108 } 109 110 if (isSwitchedOut() || inst->isSquashed()) { 111 iewStage->decrWb(inst->seqNum); 112 } else { 113 if (!state->noWB) { 114 if (!TheISA::HasUnalignedMemAcc || !state->isSplit || 115 !state->isLoad) { 116 writeback(inst, pkt); 117 } else { 118 writeback(inst, state->mainPkt); 119 } 120 } 121 122 if (inst->isStore()) { 123 completeStore(state->idx); 124 } 125 } 126 127 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) { 128 delete state->mainPkt->req; 129 delete state->mainPkt; 130 } 131 delete state; 132 delete pkt->req; 133 delete pkt; 134} 135 136template <class Impl> 137LSQUnit<Impl>::LSQUnit() 138 : loads(0), stores(0), storesToWB(0), stalled(false), 139 isStoreBlocked(false), isLoadBlocked(false), 140 loadBlockedHandled(false), hasPendingPkt(false) 141{ 142} 143 144template<class Impl> 145void 146LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params, 147 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries, 148 unsigned id) 149{ 150 cpu = cpu_ptr; 151 iewStage = iew_ptr; 152 153 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 154 155 switchedOut = false; 156 157 lsq = lsq_ptr; 158 159 lsqID = id; 160 161 // Add 1 for the sentinel entry (they are circular queues). 162 LQEntries = maxLQEntries + 1; 163 SQEntries = maxSQEntries + 1; 164 165 loadQueue.resize(LQEntries); 166 storeQueue.resize(SQEntries); 167 168 depCheckShift = params->LSQDepCheckShift; 169 checkLoads = params->LSQCheckLoads; 170 171 loadHead = loadTail = 0; 172 173 storeHead = storeWBIdx = storeTail = 0; 174 175 usedPorts = 0; 176 cachePorts = params->cachePorts; 177 178 retryPkt = NULL; 179 memDepViolator = NULL; 180 181 blockedLoadSeqNum = 0; 182} 183 184template<class Impl> 185std::string 186LSQUnit<Impl>::name() const 187{ 188 if (Impl::MaxThreads == 1) { 189 return iewStage->name() + ".lsq"; 190 } else {
| 1/* 2 * Copyright (c) 2010 ARM Limited 3 * All rights reserved 4 * 5 * The license below extends only to copyright in the software and shall 6 * not be construed as granting a license to any other intellectual 7 * property including but not limited to intellectual property relating 8 * to a hardware implementation of the functionality of the software 9 * licensed hereunder. You may use the software subject to the license 10 * terms below provided that you ensure that this notice is replicated 11 * unmodified and in its entirety in all distributions of the software, 12 * modified or unmodified, in source code or in binary form. 13 * 14 * Copyright (c) 2004-2005 The Regents of The University of Michigan 15 * All rights reserved. 16 * 17 * Redistribution and use in source and binary forms, with or without 18 * modification, are permitted provided that the following conditions are 19 * met: redistributions of source code must retain the above copyright 20 * notice, this list of conditions and the following disclaimer; 21 * redistributions in binary form must reproduce the above copyright 22 * notice, this list of conditions and the following disclaimer in the 23 * documentation and/or other materials provided with the distribution; 24 * neither the name of the copyright holders nor the names of its 25 * contributors may be used to endorse or promote products derived from 26 * this software without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Authors: Kevin Lim 41 * Korey Sewell 42 */ 43 44#include "arch/locked_mem.hh" 45#include "base/str.hh" 46#include "config/the_isa.hh" 47#include "config/use_checker.hh" 48#include "cpu/o3/lsq.hh" 49#include "cpu/o3/lsq_unit.hh" 50#include "debug/Activity.hh" 51#include "debug/IEW.hh" 52#include "debug/LSQUnit.hh" 53#include "mem/packet.hh" 54#include "mem/request.hh" 55 56#if USE_CHECKER 57#include "cpu/checker/cpu.hh" 58#endif 59 60template<class Impl> 61LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt, 62 LSQUnit *lsq_ptr) 63 : inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 64{ 65 this->setFlags(Event::AutoDelete); 66} 67 68template<class Impl> 69void 70LSQUnit<Impl>::WritebackEvent::process() 71{ 72 if (!lsqPtr->isSwitchedOut()) { 73 lsqPtr->writeback(inst, pkt); 74 } 75 76 if (pkt->senderState) 77 delete pkt->senderState; 78 79 delete pkt->req; 80 delete pkt; 81} 82 83template<class Impl> 84const char * 85LSQUnit<Impl>::WritebackEvent::description() const 86{ 87 return "Store writeback"; 88} 89 90template<class Impl> 91void 92LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 93{ 94 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 95 DynInstPtr inst = state->inst; 96 DPRINTF(IEW, "Writeback event [sn:%lli].\n", inst->seqNum); 97 DPRINTF(Activity, "Activity: Writeback event [sn:%lli].\n", inst->seqNum); 98 99 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum); 100 101 assert(!pkt->wasNacked()); 102 103 // If this is a split access, wait until all packets are received. 104 if (TheISA::HasUnalignedMemAcc && !state->complete()) { 105 delete pkt->req; 106 delete pkt; 107 return; 108 } 109 110 if (isSwitchedOut() || inst->isSquashed()) { 111 iewStage->decrWb(inst->seqNum); 112 } else { 113 if (!state->noWB) { 114 if (!TheISA::HasUnalignedMemAcc || !state->isSplit || 115 !state->isLoad) { 116 writeback(inst, pkt); 117 } else { 118 writeback(inst, state->mainPkt); 119 } 120 } 121 122 if (inst->isStore()) { 123 completeStore(state->idx); 124 } 125 } 126 127 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) { 128 delete state->mainPkt->req; 129 delete state->mainPkt; 130 } 131 delete state; 132 delete pkt->req; 133 delete pkt; 134} 135 136template <class Impl> 137LSQUnit<Impl>::LSQUnit() 138 : loads(0), stores(0), storesToWB(0), stalled(false), 139 isStoreBlocked(false), isLoadBlocked(false), 140 loadBlockedHandled(false), hasPendingPkt(false) 141{ 142} 143 144template<class Impl> 145void 146LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params, 147 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries, 148 unsigned id) 149{ 150 cpu = cpu_ptr; 151 iewStage = iew_ptr; 152 153 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 154 155 switchedOut = false; 156 157 lsq = lsq_ptr; 158 159 lsqID = id; 160 161 // Add 1 for the sentinel entry (they are circular queues). 162 LQEntries = maxLQEntries + 1; 163 SQEntries = maxSQEntries + 1; 164 165 loadQueue.resize(LQEntries); 166 storeQueue.resize(SQEntries); 167 168 depCheckShift = params->LSQDepCheckShift; 169 checkLoads = params->LSQCheckLoads; 170 171 loadHead = loadTail = 0; 172 173 storeHead = storeWBIdx = storeTail = 0; 174 175 usedPorts = 0; 176 cachePorts = params->cachePorts; 177 178 retryPkt = NULL; 179 memDepViolator = NULL; 180 181 blockedLoadSeqNum = 0; 182} 183 184template<class Impl> 185std::string 186LSQUnit<Impl>::name() const 187{ 188 if (Impl::MaxThreads == 1) { 189 return iewStage->name() + ".lsq"; 190 } else {
|
192 } 193} 194 195template<class Impl> 196void 197LSQUnit<Impl>::regStats() 198{ 199 lsqForwLoads 200 .name(name() + ".forwLoads") 201 .desc("Number of loads that had data forwarded from stores"); 202 203 invAddrLoads 204 .name(name() + ".invAddrLoads") 205 .desc("Number of loads ignored due to an invalid address"); 206 207 lsqSquashedLoads 208 .name(name() + ".squashedLoads") 209 .desc("Number of loads squashed"); 210 211 lsqIgnoredResponses 212 .name(name() + ".ignoredResponses") 213 .desc("Number of memory responses ignored because the instruction is squashed"); 214 215 lsqMemOrderViolation 216 .name(name() + ".memOrderViolation") 217 .desc("Number of memory ordering violations"); 218 219 lsqSquashedStores 220 .name(name() + ".squashedStores") 221 .desc("Number of stores squashed"); 222 223 invAddrSwpfs 224 .name(name() + ".invAddrSwpfs") 225 .desc("Number of software prefetches ignored due to an invalid address"); 226 227 lsqBlockedLoads 228 .name(name() + ".blockedLoads") 229 .desc("Number of blocked loads due to partial load-store forwarding"); 230 231 lsqRescheduledLoads 232 .name(name() + ".rescheduledLoads") 233 .desc("Number of loads that were rescheduled"); 234 235 lsqCacheBlocked 236 .name(name() + ".cacheBlocked") 237 .desc("Number of times an access to memory failed due to the cache being blocked"); 238} 239 240template<class Impl> 241void 242LSQUnit<Impl>::setDcachePort(Port *dcache_port) 243{ 244 dcachePort = dcache_port; 245 246#if USE_CHECKER 247 if (cpu->checker) { 248 cpu->checker->setDcachePort(dcachePort); 249 } 250#endif 251} 252 253template<class Impl> 254void 255LSQUnit<Impl>::clearLQ() 256{ 257 loadQueue.clear(); 258} 259 260template<class Impl> 261void 262LSQUnit<Impl>::clearSQ() 263{ 264 storeQueue.clear(); 265} 266 267template<class Impl> 268void 269LSQUnit<Impl>::switchOut() 270{ 271 switchedOut = true; 272 for (int i = 0; i < loadQueue.size(); ++i) { 273 assert(!loadQueue[i]); 274 loadQueue[i] = NULL; 275 } 276 277 assert(storesToWB == 0); 278} 279 280template<class Impl> 281void 282LSQUnit<Impl>::takeOverFrom() 283{ 284 switchedOut = false; 285 loads = stores = storesToWB = 0; 286 287 loadHead = loadTail = 0; 288 289 storeHead = storeWBIdx = storeTail = 0; 290 291 usedPorts = 0; 292 293 memDepViolator = NULL; 294 295 blockedLoadSeqNum = 0; 296 297 stalled = false; 298 isLoadBlocked = false; 299 loadBlockedHandled = false; 300} 301 302template<class Impl> 303void 304LSQUnit<Impl>::resizeLQ(unsigned size) 305{ 306 unsigned size_plus_sentinel = size + 1; 307 assert(size_plus_sentinel >= LQEntries); 308 309 if (size_plus_sentinel > LQEntries) { 310 while (size_plus_sentinel > loadQueue.size()) { 311 DynInstPtr dummy; 312 loadQueue.push_back(dummy); 313 LQEntries++; 314 } 315 } else { 316 LQEntries = size_plus_sentinel; 317 } 318 319} 320 321template<class Impl> 322void 323LSQUnit<Impl>::resizeSQ(unsigned size) 324{ 325 unsigned size_plus_sentinel = size + 1; 326 if (size_plus_sentinel > SQEntries) { 327 while (size_plus_sentinel > storeQueue.size()) { 328 SQEntry dummy; 329 storeQueue.push_back(dummy); 330 SQEntries++; 331 } 332 } else { 333 SQEntries = size_plus_sentinel; 334 } 335} 336 337template <class Impl> 338void 339LSQUnit<Impl>::insert(DynInstPtr &inst) 340{ 341 assert(inst->isMemRef()); 342 343 assert(inst->isLoad() || inst->isStore()); 344 345 if (inst->isLoad()) { 346 insertLoad(inst); 347 } else { 348 insertStore(inst); 349 } 350 351 inst->setInLSQ(); 352} 353 354template <class Impl> 355void 356LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst) 357{ 358 assert((loadTail + 1) % LQEntries != loadHead); 359 assert(loads < LQEntries); 360 361 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 362 load_inst->pcState(), loadTail, load_inst->seqNum); 363 364 load_inst->lqIdx = loadTail; 365 366 if (stores == 0) { 367 load_inst->sqIdx = -1; 368 } else { 369 load_inst->sqIdx = storeTail; 370 } 371 372 loadQueue[loadTail] = load_inst; 373 374 incrLdIdx(loadTail); 375 376 ++loads; 377} 378 379template <class Impl> 380void 381LSQUnit<Impl>::insertStore(DynInstPtr &store_inst) 382{ 383 // Make sure it is not full before inserting an instruction. 384 assert((storeTail + 1) % SQEntries != storeHead); 385 assert(stores < SQEntries); 386 387 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 388 store_inst->pcState(), storeTail, store_inst->seqNum); 389 390 store_inst->sqIdx = storeTail; 391 store_inst->lqIdx = loadTail; 392 393 storeQueue[storeTail] = SQEntry(store_inst); 394 395 incrStIdx(storeTail); 396 397 ++stores; 398} 399 400template <class Impl> 401typename Impl::DynInstPtr 402LSQUnit<Impl>::getMemDepViolator() 403{ 404 DynInstPtr temp = memDepViolator; 405 406 memDepViolator = NULL; 407 408 return temp; 409} 410 411template <class Impl> 412unsigned 413LSQUnit<Impl>::numFreeEntries() 414{ 415 unsigned free_lq_entries = LQEntries - loads; 416 unsigned free_sq_entries = SQEntries - stores; 417 418 // Both the LQ and SQ entries have an extra dummy entry to differentiate 419 // empty/full conditions. Subtract 1 from the free entries. 420 if (free_lq_entries < free_sq_entries) { 421 return free_lq_entries - 1; 422 } else { 423 return free_sq_entries - 1; 424 } 425} 426 427template <class Impl> 428int 429LSQUnit<Impl>::numLoadsReady() 430{ 431 int load_idx = loadHead; 432 int retval = 0; 433 434 while (load_idx != loadTail) { 435 assert(loadQueue[load_idx]); 436 437 if (loadQueue[load_idx]->readyToIssue()) { 438 ++retval; 439 } 440 } 441 442 return retval; 443} 444 445template <class Impl> 446Fault 447LSQUnit<Impl>::checkViolations(int load_idx, DynInstPtr &inst) 448{ 449 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 450 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 451 452 /** @todo in theory you only need to check an instruction that has executed 453 * however, there isn't a good way in the pipeline at the moment to check 454 * all instructions that will execute before the store writes back. Thus, 455 * like the implementation that came before it, we're overly conservative. 456 */ 457 while (load_idx != loadTail) { 458 DynInstPtr ld_inst = loadQueue[load_idx]; 459 if (!ld_inst->effAddrValid || ld_inst->uncacheable()) { 460 incrLdIdx(load_idx); 461 continue; 462 } 463 464 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 465 Addr ld_eff_addr2 = 466 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 467 468 if ((inst_eff_addr2 > ld_eff_addr1 && inst_eff_addr1 < ld_eff_addr2) || 469 inst_eff_addr1 == ld_eff_addr1) { 470 // A load/store incorrectly passed this load/store. 471 // Check if we already have a violator, or if it's newer 472 // squash and refetch. 473 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 474 break; 475 476 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and [sn:%lli]" 477 " at address %#x\n", inst->seqNum, ld_inst->seqNum, 478 ld_eff_addr1); 479 memDepViolator = ld_inst; 480 481 ++lsqMemOrderViolation; 482 483 return TheISA::genMachineCheckFault(); 484 } 485 486 incrLdIdx(load_idx); 487 } 488 return NoFault; 489} 490 491 492 493 494template <class Impl> 495Fault 496LSQUnit<Impl>::executeLoad(DynInstPtr &inst) 497{ 498 using namespace TheISA; 499 // Execute a specific load. 500 Fault load_fault = NoFault; 501 502 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 503 inst->pcState(), inst->seqNum); 504 505 assert(!inst->isSquashed()); 506 507 load_fault = inst->initiateAcc(); 508 509 if (inst->isTranslationDelayed() && 510 load_fault == NoFault) 511 return load_fault; 512 513 // If the instruction faulted or predicated false, then we need to send it 514 // along to commit without the instruction completing. 515 if (load_fault != NoFault || inst->readPredicate() == false) { 516 // Send this instruction to commit, also make sure iew stage 517 // realizes there is activity. 518 // Mark it as executed unless it is an uncached load that 519 // needs to hit the head of commit. 520 if (inst->readPredicate() == false) 521 inst->forwardOldRegs(); 522 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 523 inst->seqNum, 524 (load_fault != NoFault ? "fault" : "predication")); 525 if (!(inst->hasRequest() && inst->uncacheable()) || 526 inst->isAtCommit()) { 527 inst->setExecuted(); 528 } 529 iewStage->instToCommit(inst); 530 iewStage->activityThisCycle(); 531 } else if (!loadBlocked()) { 532 assert(inst->effAddrValid); 533 int load_idx = inst->lqIdx; 534 incrLdIdx(load_idx); 535 536 if (checkLoads) 537 return checkViolations(load_idx, inst); 538 } 539 540 return load_fault; 541} 542 543template <class Impl> 544Fault 545LSQUnit<Impl>::executeStore(DynInstPtr &store_inst) 546{ 547 using namespace TheISA; 548 // Make sure that a store exists. 549 assert(stores != 0); 550 551 int store_idx = store_inst->sqIdx; 552 553 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 554 store_inst->pcState(), store_inst->seqNum); 555 556 assert(!store_inst->isSquashed()); 557 558 // Check the recently completed loads to see if any match this store's 559 // address. If so, then we have a memory ordering violation. 560 int load_idx = store_inst->lqIdx; 561 562 Fault store_fault = store_inst->initiateAcc(); 563 564 if (store_inst->isTranslationDelayed() && 565 store_fault == NoFault) 566 return store_fault; 567 568 if (store_inst->readPredicate() == false) 569 store_inst->forwardOldRegs(); 570 571 if (storeQueue[store_idx].size == 0) { 572 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 573 store_inst->pcState(), store_inst->seqNum); 574 575 return store_fault; 576 } else if (store_inst->readPredicate() == false) { 577 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 578 store_inst->seqNum); 579 return store_fault; 580 } 581 582 assert(store_fault == NoFault); 583 584 if (store_inst->isStoreConditional()) { 585 // Store conditionals need to set themselves as able to 586 // writeback if we haven't had a fault by here. 587 storeQueue[store_idx].canWB = true; 588 589 ++storesToWB; 590 } 591 592 return checkViolations(load_idx, store_inst); 593 594} 595 596template <class Impl> 597void 598LSQUnit<Impl>::commitLoad() 599{ 600 assert(loadQueue[loadHead]); 601 602 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 603 loadQueue[loadHead]->pcState()); 604 605 loadQueue[loadHead] = NULL; 606 607 incrLdIdx(loadHead); 608 609 --loads; 610} 611 612template <class Impl> 613void 614LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 615{ 616 assert(loads == 0 || loadQueue[loadHead]); 617 618 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 619 commitLoad(); 620 } 621} 622 623template <class Impl> 624void 625LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 626{ 627 assert(stores == 0 || storeQueue[storeHead].inst); 628 629 int store_idx = storeHead; 630 631 while (store_idx != storeTail) { 632 assert(storeQueue[store_idx].inst); 633 // Mark any stores that are now committed and have not yet 634 // been marked as able to write back. 635 if (!storeQueue[store_idx].canWB) { 636 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 637 break; 638 } 639 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 640 "%s [sn:%lli]\n", 641 storeQueue[store_idx].inst->pcState(), 642 storeQueue[store_idx].inst->seqNum); 643 644 storeQueue[store_idx].canWB = true; 645 646 ++storesToWB; 647 } 648 649 incrStIdx(store_idx); 650 } 651} 652 653template <class Impl> 654void 655LSQUnit<Impl>::writebackPendingStore() 656{ 657 if (hasPendingPkt) { 658 assert(pendingPkt != NULL); 659 660 // If the cache is blocked, this will store the packet for retry. 661 if (sendStore(pendingPkt)) { 662 storePostSend(pendingPkt); 663 } 664 pendingPkt = NULL; 665 hasPendingPkt = false; 666 } 667} 668 669template <class Impl> 670void 671LSQUnit<Impl>::writebackStores() 672{ 673 // First writeback the second packet from any split store that didn't 674 // complete last cycle because there weren't enough cache ports available. 675 if (TheISA::HasUnalignedMemAcc) { 676 writebackPendingStore(); 677 } 678 679 while (storesToWB > 0 && 680 storeWBIdx != storeTail && 681 storeQueue[storeWBIdx].inst && 682 storeQueue[storeWBIdx].canWB && 683 usedPorts < cachePorts) { 684 685 if (isStoreBlocked || lsq->cacheBlocked()) { 686 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 687 " is blocked!\n"); 688 break; 689 } 690 691 // Store didn't write any data so no need to write it back to 692 // memory. 693 if (storeQueue[storeWBIdx].size == 0) { 694 completeStore(storeWBIdx); 695 696 incrStIdx(storeWBIdx); 697 698 continue; 699 } 700 701 ++usedPorts; 702 703 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 704 incrStIdx(storeWBIdx); 705 706 continue; 707 } 708 709 assert(storeQueue[storeWBIdx].req); 710 assert(!storeQueue[storeWBIdx].committed); 711 712 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 713 assert(storeQueue[storeWBIdx].sreqLow); 714 assert(storeQueue[storeWBIdx].sreqHigh); 715 } 716 717 DynInstPtr inst = storeQueue[storeWBIdx].inst; 718 719 Request *req = storeQueue[storeWBIdx].req; 720 storeQueue[storeWBIdx].committed = true; 721 722 assert(!inst->memData); 723 inst->memData = new uint8_t[64]; 724 725 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); 726 727 MemCmd command = 728 req->isSwap() ? MemCmd::SwapReq : 729 (req->isLLSC() ? MemCmd::StoreCondReq : MemCmd::WriteReq); 730 PacketPtr data_pkt; 731 PacketPtr snd_data_pkt = NULL; 732 733 LSQSenderState *state = new LSQSenderState; 734 state->isLoad = false; 735 state->idx = storeWBIdx; 736 state->inst = inst; 737 738 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) { 739 740 // Build a single data packet if the store isn't split. 741 data_pkt = new Packet(req, command, Packet::Broadcast); 742 data_pkt->dataStatic(inst->memData); 743 data_pkt->senderState = state; 744 } else { 745 RequestPtr sreqLow = storeQueue[storeWBIdx].sreqLow; 746 RequestPtr sreqHigh = storeQueue[storeWBIdx].sreqHigh; 747 748 // Create two packets if the store is split in two. 749 data_pkt = new Packet(sreqLow, command, Packet::Broadcast); 750 snd_data_pkt = new Packet(sreqHigh, command, Packet::Broadcast); 751 752 data_pkt->dataStatic(inst->memData); 753 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize()); 754 755 data_pkt->senderState = state; 756 snd_data_pkt->senderState = state; 757 758 state->isSplit = true; 759 state->outstanding = 2; 760 761 // Can delete the main request now. 762 delete req; 763 req = sreqLow; 764 } 765 766 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 767 "to Addr:%#x, data:%#x [sn:%lli]\n", 768 storeWBIdx, inst->pcState(), 769 req->getPaddr(), (int)*(inst->memData), 770 inst->seqNum); 771 772 // @todo: Remove this SC hack once the memory system handles it. 773 if (inst->isStoreConditional()) { 774 assert(!storeQueue[storeWBIdx].isSplit); 775 // Disable recording the result temporarily. Writing to 776 // misc regs normally updates the result, but this is not 777 // the desired behavior when handling store conditionals. 778 inst->recordResult = false; 779 bool success = TheISA::handleLockedWrite(inst.get(), req); 780 inst->recordResult = true; 781 782 if (!success) { 783 // Instantly complete this store. 784 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 785 "Instantly completing it.\n", 786 inst->seqNum); 787 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this); 788 cpu->schedule(wb, curTick() + 1); 789 completeStore(storeWBIdx); 790 incrStIdx(storeWBIdx); 791 continue; 792 } 793 } else { 794 // Non-store conditionals do not need a writeback. 795 state->noWB = true; 796 } 797 798 if (!sendStore(data_pkt)) { 799 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will" 800 "retry later\n", 801 inst->seqNum); 802 803 // Need to store the second packet, if split. 804 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 805 state->pktToSend = true; 806 state->pendingPacket = snd_data_pkt; 807 } 808 } else { 809 810 // If split, try to send the second packet too 811 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 812 assert(snd_data_pkt); 813 814 // Ensure there are enough ports to use. 815 if (usedPorts < cachePorts) { 816 ++usedPorts; 817 if (sendStore(snd_data_pkt)) { 818 storePostSend(snd_data_pkt); 819 } else { 820 DPRINTF(IEW, "D-Cache became blocked when writing" 821 " [sn:%lli] second packet, will retry later\n", 822 inst->seqNum); 823 } 824 } else { 825 826 // Store the packet for when there's free ports. 827 assert(pendingPkt == NULL); 828 pendingPkt = snd_data_pkt; 829 hasPendingPkt = true; 830 } 831 } else { 832 833 // Not a split store. 834 storePostSend(data_pkt); 835 } 836 } 837 } 838 839 // Not sure this should set it to 0. 840 usedPorts = 0; 841 842 assert(stores >= 0 && storesToWB >= 0); 843} 844 845/*template <class Impl> 846void 847LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 848{ 849 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 850 mshrSeqNums.end(), 851 seqNum); 852 853 if (mshr_it != mshrSeqNums.end()) { 854 mshrSeqNums.erase(mshr_it); 855 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 856 } 857}*/ 858 859template <class Impl> 860void 861LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 862{ 863 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 864 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 865 866 int load_idx = loadTail; 867 decrLdIdx(load_idx); 868 869 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 870 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 871 "[sn:%lli]\n", 872 loadQueue[load_idx]->pcState(), 873 loadQueue[load_idx]->seqNum); 874 875 if (isStalled() && load_idx == stallingLoadIdx) { 876 stalled = false; 877 stallingStoreIsn = 0; 878 stallingLoadIdx = 0; 879 } 880 881 // Clear the smart pointer to make sure it is decremented. 882 loadQueue[load_idx]->setSquashed(); 883 loadQueue[load_idx] = NULL; 884 --loads; 885 886 // Inefficient! 887 loadTail = load_idx; 888 889 decrLdIdx(load_idx); 890 ++lsqSquashedLoads; 891 } 892 893 if (isLoadBlocked) { 894 if (squashed_num < blockedLoadSeqNum) { 895 isLoadBlocked = false; 896 loadBlockedHandled = false; 897 blockedLoadSeqNum = 0; 898 } 899 } 900 901 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 902 memDepViolator = NULL; 903 } 904 905 int store_idx = storeTail; 906 decrStIdx(store_idx); 907 908 while (stores != 0 && 909 storeQueue[store_idx].inst->seqNum > squashed_num) { 910 // Instructions marked as can WB are already committed. 911 if (storeQueue[store_idx].canWB) { 912 break; 913 } 914 915 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 916 "idx:%i [sn:%lli]\n", 917 storeQueue[store_idx].inst->pcState(), 918 store_idx, storeQueue[store_idx].inst->seqNum); 919 920 // I don't think this can happen. It should have been cleared 921 // by the stalling load. 922 if (isStalled() && 923 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 924 panic("Is stalled should have been cleared by stalling load!\n"); 925 stalled = false; 926 stallingStoreIsn = 0; 927 } 928 929 // Clear the smart pointer to make sure it is decremented. 930 storeQueue[store_idx].inst->setSquashed(); 931 storeQueue[store_idx].inst = NULL; 932 storeQueue[store_idx].canWB = 0; 933 934 // Must delete request now that it wasn't handed off to 935 // memory. This is quite ugly. @todo: Figure out the proper 936 // place to really handle request deletes. 937 delete storeQueue[store_idx].req; 938 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) { 939 delete storeQueue[store_idx].sreqLow; 940 delete storeQueue[store_idx].sreqHigh; 941 942 storeQueue[store_idx].sreqLow = NULL; 943 storeQueue[store_idx].sreqHigh = NULL; 944 } 945 946 storeQueue[store_idx].req = NULL; 947 --stores; 948 949 // Inefficient! 950 storeTail = store_idx; 951 952 decrStIdx(store_idx); 953 ++lsqSquashedStores; 954 } 955} 956 957template <class Impl> 958void 959LSQUnit<Impl>::storePostSend(PacketPtr pkt) 960{ 961 if (isStalled() && 962 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 963 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 964 "load idx:%i\n", 965 stallingStoreIsn, stallingLoadIdx); 966 stalled = false; 967 stallingStoreIsn = 0; 968 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 969 } 970 971 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 972 // The store is basically completed at this time. This 973 // only works so long as the checker doesn't try to 974 // verify the value in memory for stores. 975 storeQueue[storeWBIdx].inst->setCompleted(); 976#if USE_CHECKER 977 if (cpu->checker) { 978 cpu->checker->verify(storeQueue[storeWBIdx].inst); 979 } 980#endif 981 } 982 983 incrStIdx(storeWBIdx); 984} 985 986template <class Impl> 987void 988LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt) 989{ 990 iewStage->wakeCPU(); 991 992 // Squashed instructions do not need to complete their access. 993 if (inst->isSquashed()) { 994 iewStage->decrWb(inst->seqNum); 995 assert(!inst->isStore()); 996 ++lsqIgnoredResponses; 997 return; 998 } 999 1000 if (!inst->isExecuted()) { 1001 inst->setExecuted(); 1002 1003 // Complete access to copy data to proper place. 1004 inst->completeAcc(pkt); 1005 } 1006 1007 // Need to insert instruction into queue to commit 1008 iewStage->instToCommit(inst); 1009 1010 iewStage->activityThisCycle(); 1011 1012 // see if this load changed the PC 1013 iewStage->checkMisprediction(inst); 1014} 1015 1016template <class Impl> 1017void 1018LSQUnit<Impl>::completeStore(int store_idx) 1019{ 1020 assert(storeQueue[store_idx].inst); 1021 storeQueue[store_idx].completed = true; 1022 --storesToWB; 1023 // A bit conservative because a store completion may not free up entries, 1024 // but hopefully avoids two store completions in one cycle from making 1025 // the CPU tick twice. 1026 cpu->wakeCPU(); 1027 cpu->activityThisCycle(); 1028 1029 if (store_idx == storeHead) { 1030 do { 1031 incrStIdx(storeHead); 1032 1033 --stores; 1034 } while (storeQueue[storeHead].completed && 1035 storeHead != storeTail); 1036 1037 iewStage->updateLSQNextCycle = true; 1038 } 1039 1040 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1041 "idx:%i\n", 1042 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 1043 1044 if (isStalled() && 1045 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1046 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1047 "load idx:%i\n", 1048 stallingStoreIsn, stallingLoadIdx); 1049 stalled = false; 1050 stallingStoreIsn = 0; 1051 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1052 } 1053 1054 storeQueue[store_idx].inst->setCompleted(); 1055 1056 // Tell the checker we've completed this instruction. Some stores 1057 // may get reported twice to the checker, but the checker can 1058 // handle that case. 1059#if USE_CHECKER 1060 if (cpu->checker) { 1061 cpu->checker->verify(storeQueue[store_idx].inst); 1062 } 1063#endif 1064} 1065 1066template <class Impl> 1067bool 1068LSQUnit<Impl>::sendStore(PacketPtr data_pkt) 1069{ 1070 if (!dcachePort->sendTiming(data_pkt)) { 1071 // Need to handle becoming blocked on a store. 1072 isStoreBlocked = true; 1073 ++lsqCacheBlocked; 1074 assert(retryPkt == NULL); 1075 retryPkt = data_pkt; 1076 lsq->setRetryTid(lsqID); 1077 return false; 1078 } 1079 return true; 1080} 1081 1082template <class Impl> 1083void 1084LSQUnit<Impl>::recvRetry() 1085{ 1086 if (isStoreBlocked) { 1087 DPRINTF(LSQUnit, "Receiving retry: store blocked\n"); 1088 assert(retryPkt != NULL); 1089 1090 if (dcachePort->sendTiming(retryPkt)) { 1091 LSQSenderState *state = 1092 dynamic_cast<LSQSenderState *>(retryPkt->senderState); 1093 1094 // Don't finish the store unless this is the last packet. 1095 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend || 1096 state->pendingPacket == retryPkt) { 1097 state->pktToSend = false; 1098 storePostSend(retryPkt); 1099 } 1100 retryPkt = NULL; 1101 isStoreBlocked = false; 1102 lsq->setRetryTid(InvalidThreadID); 1103 1104 // Send any outstanding packet. 1105 if (TheISA::HasUnalignedMemAcc && state->pktToSend) { 1106 assert(state->pendingPacket); 1107 if (sendStore(state->pendingPacket)) { 1108 storePostSend(state->pendingPacket); 1109 } 1110 } 1111 } else { 1112 // Still blocked! 1113 ++lsqCacheBlocked; 1114 lsq->setRetryTid(lsqID); 1115 } 1116 } else if (isLoadBlocked) { 1117 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, " 1118 "no need to resend packet.\n"); 1119 } else { 1120 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n"); 1121 } 1122} 1123 1124template <class Impl> 1125inline void 1126LSQUnit<Impl>::incrStIdx(int &store_idx) 1127{ 1128 if (++store_idx >= SQEntries) 1129 store_idx = 0; 1130} 1131 1132template <class Impl> 1133inline void 1134LSQUnit<Impl>::decrStIdx(int &store_idx) 1135{ 1136 if (--store_idx < 0) 1137 store_idx += SQEntries; 1138} 1139 1140template <class Impl> 1141inline void 1142LSQUnit<Impl>::incrLdIdx(int &load_idx) 1143{ 1144 if (++load_idx >= LQEntries) 1145 load_idx = 0; 1146} 1147 1148template <class Impl> 1149inline void 1150LSQUnit<Impl>::decrLdIdx(int &load_idx) 1151{ 1152 if (--load_idx < 0) 1153 load_idx += LQEntries; 1154} 1155 1156template <class Impl> 1157void 1158LSQUnit<Impl>::dumpInsts() 1159{ 1160 cprintf("Load store queue: Dumping instructions.\n"); 1161 cprintf("Load queue size: %i\n", loads); 1162 cprintf("Load queue: "); 1163 1164 int load_idx = loadHead; 1165 1166 while (load_idx != loadTail && loadQueue[load_idx]) { 1167 cprintf("%s ", loadQueue[load_idx]->pcState()); 1168 1169 incrLdIdx(load_idx); 1170 } 1171 1172 cprintf("Store queue size: %i\n", stores); 1173 cprintf("Store queue: "); 1174 1175 int store_idx = storeHead; 1176 1177 while (store_idx != storeTail && storeQueue[store_idx].inst) { 1178 cprintf("%s ", storeQueue[store_idx].inst->pcState()); 1179 1180 incrStIdx(store_idx); 1181 } 1182 1183 cprintf("\n"); 1184}
| 192 } 193} 194 195template<class Impl> 196void 197LSQUnit<Impl>::regStats() 198{ 199 lsqForwLoads 200 .name(name() + ".forwLoads") 201 .desc("Number of loads that had data forwarded from stores"); 202 203 invAddrLoads 204 .name(name() + ".invAddrLoads") 205 .desc("Number of loads ignored due to an invalid address"); 206 207 lsqSquashedLoads 208 .name(name() + ".squashedLoads") 209 .desc("Number of loads squashed"); 210 211 lsqIgnoredResponses 212 .name(name() + ".ignoredResponses") 213 .desc("Number of memory responses ignored because the instruction is squashed"); 214 215 lsqMemOrderViolation 216 .name(name() + ".memOrderViolation") 217 .desc("Number of memory ordering violations"); 218 219 lsqSquashedStores 220 .name(name() + ".squashedStores") 221 .desc("Number of stores squashed"); 222 223 invAddrSwpfs 224 .name(name() + ".invAddrSwpfs") 225 .desc("Number of software prefetches ignored due to an invalid address"); 226 227 lsqBlockedLoads 228 .name(name() + ".blockedLoads") 229 .desc("Number of blocked loads due to partial load-store forwarding"); 230 231 lsqRescheduledLoads 232 .name(name() + ".rescheduledLoads") 233 .desc("Number of loads that were rescheduled"); 234 235 lsqCacheBlocked 236 .name(name() + ".cacheBlocked") 237 .desc("Number of times an access to memory failed due to the cache being blocked"); 238} 239 240template<class Impl> 241void 242LSQUnit<Impl>::setDcachePort(Port *dcache_port) 243{ 244 dcachePort = dcache_port; 245 246#if USE_CHECKER 247 if (cpu->checker) { 248 cpu->checker->setDcachePort(dcachePort); 249 } 250#endif 251} 252 253template<class Impl> 254void 255LSQUnit<Impl>::clearLQ() 256{ 257 loadQueue.clear(); 258} 259 260template<class Impl> 261void 262LSQUnit<Impl>::clearSQ() 263{ 264 storeQueue.clear(); 265} 266 267template<class Impl> 268void 269LSQUnit<Impl>::switchOut() 270{ 271 switchedOut = true; 272 for (int i = 0; i < loadQueue.size(); ++i) { 273 assert(!loadQueue[i]); 274 loadQueue[i] = NULL; 275 } 276 277 assert(storesToWB == 0); 278} 279 280template<class Impl> 281void 282LSQUnit<Impl>::takeOverFrom() 283{ 284 switchedOut = false; 285 loads = stores = storesToWB = 0; 286 287 loadHead = loadTail = 0; 288 289 storeHead = storeWBIdx = storeTail = 0; 290 291 usedPorts = 0; 292 293 memDepViolator = NULL; 294 295 blockedLoadSeqNum = 0; 296 297 stalled = false; 298 isLoadBlocked = false; 299 loadBlockedHandled = false; 300} 301 302template<class Impl> 303void 304LSQUnit<Impl>::resizeLQ(unsigned size) 305{ 306 unsigned size_plus_sentinel = size + 1; 307 assert(size_plus_sentinel >= LQEntries); 308 309 if (size_plus_sentinel > LQEntries) { 310 while (size_plus_sentinel > loadQueue.size()) { 311 DynInstPtr dummy; 312 loadQueue.push_back(dummy); 313 LQEntries++; 314 } 315 } else { 316 LQEntries = size_plus_sentinel; 317 } 318 319} 320 321template<class Impl> 322void 323LSQUnit<Impl>::resizeSQ(unsigned size) 324{ 325 unsigned size_plus_sentinel = size + 1; 326 if (size_plus_sentinel > SQEntries) { 327 while (size_plus_sentinel > storeQueue.size()) { 328 SQEntry dummy; 329 storeQueue.push_back(dummy); 330 SQEntries++; 331 } 332 } else { 333 SQEntries = size_plus_sentinel; 334 } 335} 336 337template <class Impl> 338void 339LSQUnit<Impl>::insert(DynInstPtr &inst) 340{ 341 assert(inst->isMemRef()); 342 343 assert(inst->isLoad() || inst->isStore()); 344 345 if (inst->isLoad()) { 346 insertLoad(inst); 347 } else { 348 insertStore(inst); 349 } 350 351 inst->setInLSQ(); 352} 353 354template <class Impl> 355void 356LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst) 357{ 358 assert((loadTail + 1) % LQEntries != loadHead); 359 assert(loads < LQEntries); 360 361 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 362 load_inst->pcState(), loadTail, load_inst->seqNum); 363 364 load_inst->lqIdx = loadTail; 365 366 if (stores == 0) { 367 load_inst->sqIdx = -1; 368 } else { 369 load_inst->sqIdx = storeTail; 370 } 371 372 loadQueue[loadTail] = load_inst; 373 374 incrLdIdx(loadTail); 375 376 ++loads; 377} 378 379template <class Impl> 380void 381LSQUnit<Impl>::insertStore(DynInstPtr &store_inst) 382{ 383 // Make sure it is not full before inserting an instruction. 384 assert((storeTail + 1) % SQEntries != storeHead); 385 assert(stores < SQEntries); 386 387 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 388 store_inst->pcState(), storeTail, store_inst->seqNum); 389 390 store_inst->sqIdx = storeTail; 391 store_inst->lqIdx = loadTail; 392 393 storeQueue[storeTail] = SQEntry(store_inst); 394 395 incrStIdx(storeTail); 396 397 ++stores; 398} 399 400template <class Impl> 401typename Impl::DynInstPtr 402LSQUnit<Impl>::getMemDepViolator() 403{ 404 DynInstPtr temp = memDepViolator; 405 406 memDepViolator = NULL; 407 408 return temp; 409} 410 411template <class Impl> 412unsigned 413LSQUnit<Impl>::numFreeEntries() 414{ 415 unsigned free_lq_entries = LQEntries - loads; 416 unsigned free_sq_entries = SQEntries - stores; 417 418 // Both the LQ and SQ entries have an extra dummy entry to differentiate 419 // empty/full conditions. Subtract 1 from the free entries. 420 if (free_lq_entries < free_sq_entries) { 421 return free_lq_entries - 1; 422 } else { 423 return free_sq_entries - 1; 424 } 425} 426 427template <class Impl> 428int 429LSQUnit<Impl>::numLoadsReady() 430{ 431 int load_idx = loadHead; 432 int retval = 0; 433 434 while (load_idx != loadTail) { 435 assert(loadQueue[load_idx]); 436 437 if (loadQueue[load_idx]->readyToIssue()) { 438 ++retval; 439 } 440 } 441 442 return retval; 443} 444 445template <class Impl> 446Fault 447LSQUnit<Impl>::checkViolations(int load_idx, DynInstPtr &inst) 448{ 449 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 450 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 451 452 /** @todo in theory you only need to check an instruction that has executed 453 * however, there isn't a good way in the pipeline at the moment to check 454 * all instructions that will execute before the store writes back. Thus, 455 * like the implementation that came before it, we're overly conservative. 456 */ 457 while (load_idx != loadTail) { 458 DynInstPtr ld_inst = loadQueue[load_idx]; 459 if (!ld_inst->effAddrValid || ld_inst->uncacheable()) { 460 incrLdIdx(load_idx); 461 continue; 462 } 463 464 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 465 Addr ld_eff_addr2 = 466 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 467 468 if ((inst_eff_addr2 > ld_eff_addr1 && inst_eff_addr1 < ld_eff_addr2) || 469 inst_eff_addr1 == ld_eff_addr1) { 470 // A load/store incorrectly passed this load/store. 471 // Check if we already have a violator, or if it's newer 472 // squash and refetch. 473 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 474 break; 475 476 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and [sn:%lli]" 477 " at address %#x\n", inst->seqNum, ld_inst->seqNum, 478 ld_eff_addr1); 479 memDepViolator = ld_inst; 480 481 ++lsqMemOrderViolation; 482 483 return TheISA::genMachineCheckFault(); 484 } 485 486 incrLdIdx(load_idx); 487 } 488 return NoFault; 489} 490 491 492 493 494template <class Impl> 495Fault 496LSQUnit<Impl>::executeLoad(DynInstPtr &inst) 497{ 498 using namespace TheISA; 499 // Execute a specific load. 500 Fault load_fault = NoFault; 501 502 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 503 inst->pcState(), inst->seqNum); 504 505 assert(!inst->isSquashed()); 506 507 load_fault = inst->initiateAcc(); 508 509 if (inst->isTranslationDelayed() && 510 load_fault == NoFault) 511 return load_fault; 512 513 // If the instruction faulted or predicated false, then we need to send it 514 // along to commit without the instruction completing. 515 if (load_fault != NoFault || inst->readPredicate() == false) { 516 // Send this instruction to commit, also make sure iew stage 517 // realizes there is activity. 518 // Mark it as executed unless it is an uncached load that 519 // needs to hit the head of commit. 520 if (inst->readPredicate() == false) 521 inst->forwardOldRegs(); 522 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 523 inst->seqNum, 524 (load_fault != NoFault ? "fault" : "predication")); 525 if (!(inst->hasRequest() && inst->uncacheable()) || 526 inst->isAtCommit()) { 527 inst->setExecuted(); 528 } 529 iewStage->instToCommit(inst); 530 iewStage->activityThisCycle(); 531 } else if (!loadBlocked()) { 532 assert(inst->effAddrValid); 533 int load_idx = inst->lqIdx; 534 incrLdIdx(load_idx); 535 536 if (checkLoads) 537 return checkViolations(load_idx, inst); 538 } 539 540 return load_fault; 541} 542 543template <class Impl> 544Fault 545LSQUnit<Impl>::executeStore(DynInstPtr &store_inst) 546{ 547 using namespace TheISA; 548 // Make sure that a store exists. 549 assert(stores != 0); 550 551 int store_idx = store_inst->sqIdx; 552 553 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 554 store_inst->pcState(), store_inst->seqNum); 555 556 assert(!store_inst->isSquashed()); 557 558 // Check the recently completed loads to see if any match this store's 559 // address. If so, then we have a memory ordering violation. 560 int load_idx = store_inst->lqIdx; 561 562 Fault store_fault = store_inst->initiateAcc(); 563 564 if (store_inst->isTranslationDelayed() && 565 store_fault == NoFault) 566 return store_fault; 567 568 if (store_inst->readPredicate() == false) 569 store_inst->forwardOldRegs(); 570 571 if (storeQueue[store_idx].size == 0) { 572 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 573 store_inst->pcState(), store_inst->seqNum); 574 575 return store_fault; 576 } else if (store_inst->readPredicate() == false) { 577 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 578 store_inst->seqNum); 579 return store_fault; 580 } 581 582 assert(store_fault == NoFault); 583 584 if (store_inst->isStoreConditional()) { 585 // Store conditionals need to set themselves as able to 586 // writeback if we haven't had a fault by here. 587 storeQueue[store_idx].canWB = true; 588 589 ++storesToWB; 590 } 591 592 return checkViolations(load_idx, store_inst); 593 594} 595 596template <class Impl> 597void 598LSQUnit<Impl>::commitLoad() 599{ 600 assert(loadQueue[loadHead]); 601 602 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 603 loadQueue[loadHead]->pcState()); 604 605 loadQueue[loadHead] = NULL; 606 607 incrLdIdx(loadHead); 608 609 --loads; 610} 611 612template <class Impl> 613void 614LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 615{ 616 assert(loads == 0 || loadQueue[loadHead]); 617 618 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 619 commitLoad(); 620 } 621} 622 623template <class Impl> 624void 625LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 626{ 627 assert(stores == 0 || storeQueue[storeHead].inst); 628 629 int store_idx = storeHead; 630 631 while (store_idx != storeTail) { 632 assert(storeQueue[store_idx].inst); 633 // Mark any stores that are now committed and have not yet 634 // been marked as able to write back. 635 if (!storeQueue[store_idx].canWB) { 636 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 637 break; 638 } 639 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 640 "%s [sn:%lli]\n", 641 storeQueue[store_idx].inst->pcState(), 642 storeQueue[store_idx].inst->seqNum); 643 644 storeQueue[store_idx].canWB = true; 645 646 ++storesToWB; 647 } 648 649 incrStIdx(store_idx); 650 } 651} 652 653template <class Impl> 654void 655LSQUnit<Impl>::writebackPendingStore() 656{ 657 if (hasPendingPkt) { 658 assert(pendingPkt != NULL); 659 660 // If the cache is blocked, this will store the packet for retry. 661 if (sendStore(pendingPkt)) { 662 storePostSend(pendingPkt); 663 } 664 pendingPkt = NULL; 665 hasPendingPkt = false; 666 } 667} 668 669template <class Impl> 670void 671LSQUnit<Impl>::writebackStores() 672{ 673 // First writeback the second packet from any split store that didn't 674 // complete last cycle because there weren't enough cache ports available. 675 if (TheISA::HasUnalignedMemAcc) { 676 writebackPendingStore(); 677 } 678 679 while (storesToWB > 0 && 680 storeWBIdx != storeTail && 681 storeQueue[storeWBIdx].inst && 682 storeQueue[storeWBIdx].canWB && 683 usedPorts < cachePorts) { 684 685 if (isStoreBlocked || lsq->cacheBlocked()) { 686 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 687 " is blocked!\n"); 688 break; 689 } 690 691 // Store didn't write any data so no need to write it back to 692 // memory. 693 if (storeQueue[storeWBIdx].size == 0) { 694 completeStore(storeWBIdx); 695 696 incrStIdx(storeWBIdx); 697 698 continue; 699 } 700 701 ++usedPorts; 702 703 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 704 incrStIdx(storeWBIdx); 705 706 continue; 707 } 708 709 assert(storeQueue[storeWBIdx].req); 710 assert(!storeQueue[storeWBIdx].committed); 711 712 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 713 assert(storeQueue[storeWBIdx].sreqLow); 714 assert(storeQueue[storeWBIdx].sreqHigh); 715 } 716 717 DynInstPtr inst = storeQueue[storeWBIdx].inst; 718 719 Request *req = storeQueue[storeWBIdx].req; 720 storeQueue[storeWBIdx].committed = true; 721 722 assert(!inst->memData); 723 inst->memData = new uint8_t[64]; 724 725 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); 726 727 MemCmd command = 728 req->isSwap() ? MemCmd::SwapReq : 729 (req->isLLSC() ? MemCmd::StoreCondReq : MemCmd::WriteReq); 730 PacketPtr data_pkt; 731 PacketPtr snd_data_pkt = NULL; 732 733 LSQSenderState *state = new LSQSenderState; 734 state->isLoad = false; 735 state->idx = storeWBIdx; 736 state->inst = inst; 737 738 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) { 739 740 // Build a single data packet if the store isn't split. 741 data_pkt = new Packet(req, command, Packet::Broadcast); 742 data_pkt->dataStatic(inst->memData); 743 data_pkt->senderState = state; 744 } else { 745 RequestPtr sreqLow = storeQueue[storeWBIdx].sreqLow; 746 RequestPtr sreqHigh = storeQueue[storeWBIdx].sreqHigh; 747 748 // Create two packets if the store is split in two. 749 data_pkt = new Packet(sreqLow, command, Packet::Broadcast); 750 snd_data_pkt = new Packet(sreqHigh, command, Packet::Broadcast); 751 752 data_pkt->dataStatic(inst->memData); 753 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize()); 754 755 data_pkt->senderState = state; 756 snd_data_pkt->senderState = state; 757 758 state->isSplit = true; 759 state->outstanding = 2; 760 761 // Can delete the main request now. 762 delete req; 763 req = sreqLow; 764 } 765 766 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 767 "to Addr:%#x, data:%#x [sn:%lli]\n", 768 storeWBIdx, inst->pcState(), 769 req->getPaddr(), (int)*(inst->memData), 770 inst->seqNum); 771 772 // @todo: Remove this SC hack once the memory system handles it. 773 if (inst->isStoreConditional()) { 774 assert(!storeQueue[storeWBIdx].isSplit); 775 // Disable recording the result temporarily. Writing to 776 // misc regs normally updates the result, but this is not 777 // the desired behavior when handling store conditionals. 778 inst->recordResult = false; 779 bool success = TheISA::handleLockedWrite(inst.get(), req); 780 inst->recordResult = true; 781 782 if (!success) { 783 // Instantly complete this store. 784 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 785 "Instantly completing it.\n", 786 inst->seqNum); 787 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this); 788 cpu->schedule(wb, curTick() + 1); 789 completeStore(storeWBIdx); 790 incrStIdx(storeWBIdx); 791 continue; 792 } 793 } else { 794 // Non-store conditionals do not need a writeback. 795 state->noWB = true; 796 } 797 798 if (!sendStore(data_pkt)) { 799 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will" 800 "retry later\n", 801 inst->seqNum); 802 803 // Need to store the second packet, if split. 804 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 805 state->pktToSend = true; 806 state->pendingPacket = snd_data_pkt; 807 } 808 } else { 809 810 // If split, try to send the second packet too 811 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 812 assert(snd_data_pkt); 813 814 // Ensure there are enough ports to use. 815 if (usedPorts < cachePorts) { 816 ++usedPorts; 817 if (sendStore(snd_data_pkt)) { 818 storePostSend(snd_data_pkt); 819 } else { 820 DPRINTF(IEW, "D-Cache became blocked when writing" 821 " [sn:%lli] second packet, will retry later\n", 822 inst->seqNum); 823 } 824 } else { 825 826 // Store the packet for when there's free ports. 827 assert(pendingPkt == NULL); 828 pendingPkt = snd_data_pkt; 829 hasPendingPkt = true; 830 } 831 } else { 832 833 // Not a split store. 834 storePostSend(data_pkt); 835 } 836 } 837 } 838 839 // Not sure this should set it to 0. 840 usedPorts = 0; 841 842 assert(stores >= 0 && storesToWB >= 0); 843} 844 845/*template <class Impl> 846void 847LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 848{ 849 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 850 mshrSeqNums.end(), 851 seqNum); 852 853 if (mshr_it != mshrSeqNums.end()) { 854 mshrSeqNums.erase(mshr_it); 855 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 856 } 857}*/ 858 859template <class Impl> 860void 861LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 862{ 863 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 864 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 865 866 int load_idx = loadTail; 867 decrLdIdx(load_idx); 868 869 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 870 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 871 "[sn:%lli]\n", 872 loadQueue[load_idx]->pcState(), 873 loadQueue[load_idx]->seqNum); 874 875 if (isStalled() && load_idx == stallingLoadIdx) { 876 stalled = false; 877 stallingStoreIsn = 0; 878 stallingLoadIdx = 0; 879 } 880 881 // Clear the smart pointer to make sure it is decremented. 882 loadQueue[load_idx]->setSquashed(); 883 loadQueue[load_idx] = NULL; 884 --loads; 885 886 // Inefficient! 887 loadTail = load_idx; 888 889 decrLdIdx(load_idx); 890 ++lsqSquashedLoads; 891 } 892 893 if (isLoadBlocked) { 894 if (squashed_num < blockedLoadSeqNum) { 895 isLoadBlocked = false; 896 loadBlockedHandled = false; 897 blockedLoadSeqNum = 0; 898 } 899 } 900 901 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 902 memDepViolator = NULL; 903 } 904 905 int store_idx = storeTail; 906 decrStIdx(store_idx); 907 908 while (stores != 0 && 909 storeQueue[store_idx].inst->seqNum > squashed_num) { 910 // Instructions marked as can WB are already committed. 911 if (storeQueue[store_idx].canWB) { 912 break; 913 } 914 915 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 916 "idx:%i [sn:%lli]\n", 917 storeQueue[store_idx].inst->pcState(), 918 store_idx, storeQueue[store_idx].inst->seqNum); 919 920 // I don't think this can happen. It should have been cleared 921 // by the stalling load. 922 if (isStalled() && 923 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 924 panic("Is stalled should have been cleared by stalling load!\n"); 925 stalled = false; 926 stallingStoreIsn = 0; 927 } 928 929 // Clear the smart pointer to make sure it is decremented. 930 storeQueue[store_idx].inst->setSquashed(); 931 storeQueue[store_idx].inst = NULL; 932 storeQueue[store_idx].canWB = 0; 933 934 // Must delete request now that it wasn't handed off to 935 // memory. This is quite ugly. @todo: Figure out the proper 936 // place to really handle request deletes. 937 delete storeQueue[store_idx].req; 938 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) { 939 delete storeQueue[store_idx].sreqLow; 940 delete storeQueue[store_idx].sreqHigh; 941 942 storeQueue[store_idx].sreqLow = NULL; 943 storeQueue[store_idx].sreqHigh = NULL; 944 } 945 946 storeQueue[store_idx].req = NULL; 947 --stores; 948 949 // Inefficient! 950 storeTail = store_idx; 951 952 decrStIdx(store_idx); 953 ++lsqSquashedStores; 954 } 955} 956 957template <class Impl> 958void 959LSQUnit<Impl>::storePostSend(PacketPtr pkt) 960{ 961 if (isStalled() && 962 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 963 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 964 "load idx:%i\n", 965 stallingStoreIsn, stallingLoadIdx); 966 stalled = false; 967 stallingStoreIsn = 0; 968 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 969 } 970 971 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 972 // The store is basically completed at this time. This 973 // only works so long as the checker doesn't try to 974 // verify the value in memory for stores. 975 storeQueue[storeWBIdx].inst->setCompleted(); 976#if USE_CHECKER 977 if (cpu->checker) { 978 cpu->checker->verify(storeQueue[storeWBIdx].inst); 979 } 980#endif 981 } 982 983 incrStIdx(storeWBIdx); 984} 985 986template <class Impl> 987void 988LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt) 989{ 990 iewStage->wakeCPU(); 991 992 // Squashed instructions do not need to complete their access. 993 if (inst->isSquashed()) { 994 iewStage->decrWb(inst->seqNum); 995 assert(!inst->isStore()); 996 ++lsqIgnoredResponses; 997 return; 998 } 999 1000 if (!inst->isExecuted()) { 1001 inst->setExecuted(); 1002 1003 // Complete access to copy data to proper place. 1004 inst->completeAcc(pkt); 1005 } 1006 1007 // Need to insert instruction into queue to commit 1008 iewStage->instToCommit(inst); 1009 1010 iewStage->activityThisCycle(); 1011 1012 // see if this load changed the PC 1013 iewStage->checkMisprediction(inst); 1014} 1015 1016template <class Impl> 1017void 1018LSQUnit<Impl>::completeStore(int store_idx) 1019{ 1020 assert(storeQueue[store_idx].inst); 1021 storeQueue[store_idx].completed = true; 1022 --storesToWB; 1023 // A bit conservative because a store completion may not free up entries, 1024 // but hopefully avoids two store completions in one cycle from making 1025 // the CPU tick twice. 1026 cpu->wakeCPU(); 1027 cpu->activityThisCycle(); 1028 1029 if (store_idx == storeHead) { 1030 do { 1031 incrStIdx(storeHead); 1032 1033 --stores; 1034 } while (storeQueue[storeHead].completed && 1035 storeHead != storeTail); 1036 1037 iewStage->updateLSQNextCycle = true; 1038 } 1039 1040 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1041 "idx:%i\n", 1042 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 1043 1044 if (isStalled() && 1045 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1046 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1047 "load idx:%i\n", 1048 stallingStoreIsn, stallingLoadIdx); 1049 stalled = false; 1050 stallingStoreIsn = 0; 1051 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1052 } 1053 1054 storeQueue[store_idx].inst->setCompleted(); 1055 1056 // Tell the checker we've completed this instruction. Some stores 1057 // may get reported twice to the checker, but the checker can 1058 // handle that case. 1059#if USE_CHECKER 1060 if (cpu->checker) { 1061 cpu->checker->verify(storeQueue[store_idx].inst); 1062 } 1063#endif 1064} 1065 1066template <class Impl> 1067bool 1068LSQUnit<Impl>::sendStore(PacketPtr data_pkt) 1069{ 1070 if (!dcachePort->sendTiming(data_pkt)) { 1071 // Need to handle becoming blocked on a store. 1072 isStoreBlocked = true; 1073 ++lsqCacheBlocked; 1074 assert(retryPkt == NULL); 1075 retryPkt = data_pkt; 1076 lsq->setRetryTid(lsqID); 1077 return false; 1078 } 1079 return true; 1080} 1081 1082template <class Impl> 1083void 1084LSQUnit<Impl>::recvRetry() 1085{ 1086 if (isStoreBlocked) { 1087 DPRINTF(LSQUnit, "Receiving retry: store blocked\n"); 1088 assert(retryPkt != NULL); 1089 1090 if (dcachePort->sendTiming(retryPkt)) { 1091 LSQSenderState *state = 1092 dynamic_cast<LSQSenderState *>(retryPkt->senderState); 1093 1094 // Don't finish the store unless this is the last packet. 1095 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend || 1096 state->pendingPacket == retryPkt) { 1097 state->pktToSend = false; 1098 storePostSend(retryPkt); 1099 } 1100 retryPkt = NULL; 1101 isStoreBlocked = false; 1102 lsq->setRetryTid(InvalidThreadID); 1103 1104 // Send any outstanding packet. 1105 if (TheISA::HasUnalignedMemAcc && state->pktToSend) { 1106 assert(state->pendingPacket); 1107 if (sendStore(state->pendingPacket)) { 1108 storePostSend(state->pendingPacket); 1109 } 1110 } 1111 } else { 1112 // Still blocked! 1113 ++lsqCacheBlocked; 1114 lsq->setRetryTid(lsqID); 1115 } 1116 } else if (isLoadBlocked) { 1117 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, " 1118 "no need to resend packet.\n"); 1119 } else { 1120 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n"); 1121 } 1122} 1123 1124template <class Impl> 1125inline void 1126LSQUnit<Impl>::incrStIdx(int &store_idx) 1127{ 1128 if (++store_idx >= SQEntries) 1129 store_idx = 0; 1130} 1131 1132template <class Impl> 1133inline void 1134LSQUnit<Impl>::decrStIdx(int &store_idx) 1135{ 1136 if (--store_idx < 0) 1137 store_idx += SQEntries; 1138} 1139 1140template <class Impl> 1141inline void 1142LSQUnit<Impl>::incrLdIdx(int &load_idx) 1143{ 1144 if (++load_idx >= LQEntries) 1145 load_idx = 0; 1146} 1147 1148template <class Impl> 1149inline void 1150LSQUnit<Impl>::decrLdIdx(int &load_idx) 1151{ 1152 if (--load_idx < 0) 1153 load_idx += LQEntries; 1154} 1155 1156template <class Impl> 1157void 1158LSQUnit<Impl>::dumpInsts() 1159{ 1160 cprintf("Load store queue: Dumping instructions.\n"); 1161 cprintf("Load queue size: %i\n", loads); 1162 cprintf("Load queue: "); 1163 1164 int load_idx = loadHead; 1165 1166 while (load_idx != loadTail && loadQueue[load_idx]) { 1167 cprintf("%s ", loadQueue[load_idx]->pcState()); 1168 1169 incrLdIdx(load_idx); 1170 } 1171 1172 cprintf("Store queue size: %i\n", stores); 1173 cprintf("Store queue: "); 1174 1175 int store_idx = storeHead; 1176 1177 while (store_idx != storeTail && storeQueue[store_idx].inst) { 1178 cprintf("%s ", storeQueue[store_idx].inst->pcState()); 1179 1180 incrStIdx(store_idx); 1181 } 1182 1183 cprintf("\n"); 1184}
|