Deleted Added
sdiff udiff text old ( 11356:a80884911971 ) new ( 11357:6668387fa488 )
full compact
1
2/*
3 * Copyright (c) 2010-2014 ARM Limited
4 * Copyright (c) 2013 Advanced Micro Devices, Inc.
5 * All rights reserved
6 *
7 * The license below extends only to copyright in the software and shall
8 * not be construed as granting a license to any other intellectual
9 * property including but not limited to intellectual property relating
10 * to a hardware implementation of the functionality of the software
11 * licensed hereunder. You may use the software subject to the license
12 * terms below provided that you ensure that this notice is replicated
13 * unmodified and in its entirety in all distributions of the software,
14 * modified or unmodified, in source code or in binary form.
15 *
16 * Copyright (c) 2004-2005 The Regents of The University of Michigan
17 * All rights reserved.
18 *
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions are
21 * met: redistributions of source code must retain the above copyright
22 * notice, this list of conditions and the following disclaimer;
23 * redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution;
26 * neither the name of the copyright holders nor the names of its
27 * contributors may be used to endorse or promote products derived from
28 * this software without specific prior written permission.
29 *
30 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
31 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
32 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
33 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
34 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
35 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
36 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
37 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
38 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
39 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
40 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41 *
42 * Authors: Kevin Lim
43 * Korey Sewell
44 */
45
46#ifndef __CPU_O3_LSQ_UNIT_IMPL_HH__
47#define __CPU_O3_LSQ_UNIT_IMPL_HH__
48
49#include "arch/generic/debugfaults.hh"
50#include "arch/locked_mem.hh"
51#include "base/str.hh"
52#include "config/the_isa.hh"
53#include "cpu/checker/cpu.hh"
54#include "cpu/o3/lsq.hh"
55#include "cpu/o3/lsq_unit.hh"
56#include "debug/Activity.hh"
57#include "debug/IEW.hh"
58#include "debug/LSQUnit.hh"
59#include "debug/O3PipeView.hh"
60#include "mem/packet.hh"
61#include "mem/request.hh"
62
63template<class Impl>
64LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt,
65 LSQUnit *lsq_ptr)
66 : Event(Default_Pri, AutoDelete),
67 inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr)
68{
69}
70
71template<class Impl>
72void
73LSQUnit<Impl>::WritebackEvent::process()
74{
75 assert(!lsqPtr->cpu->switchedOut());
76
77 lsqPtr->writeback(inst, pkt);
78
79 if (pkt->senderState)
80 delete pkt->senderState;
81
82 delete pkt->req;
83 delete pkt;
84}
85
86template<class Impl>
87const char *
88LSQUnit<Impl>::WritebackEvent::description() const
89{
90 return "Store writeback";
91}
92
93template<class Impl>
94void
95LSQUnit<Impl>::completeDataAccess(PacketPtr pkt)
96{
97 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState);
98 DynInstPtr inst = state->inst;
99 DPRINTF(IEW, "Writeback event [sn:%lli].\n", inst->seqNum);
100 DPRINTF(Activity, "Activity: Writeback event [sn:%lli].\n", inst->seqNum);
101
102 if (state->cacheBlocked) {
103 // This is the first half of a previous split load,
104 // where the 2nd half blocked, ignore this response
105 DPRINTF(IEW, "[sn:%lli]: Response from first half of earlier "
106 "blocked split load recieved. Ignoring.\n", inst->seqNum);
107 delete state;
108 return;
109 }
110
111 // If this is a split access, wait until all packets are received.
112 if (TheISA::HasUnalignedMemAcc && !state->complete()) {
113 return;
114 }
115
116 assert(!cpu->switchedOut());
117 if (!inst->isSquashed()) {
118 if (!state->noWB) {
119 if (!TheISA::HasUnalignedMemAcc || !state->isSplit ||
120 !state->isLoad) {
121 writeback(inst, pkt);
122 } else {
123 writeback(inst, state->mainPkt);
124 }
125 }
126
127 if (inst->isStore()) {
128 completeStore(state->idx);
129 }
130 }
131
132 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) {
133 delete state->mainPkt->req;
134 delete state->mainPkt;
135 }
136
137 pkt->req->setAccessLatency();
138 cpu->ppDataAccessComplete->notify(std::make_pair(inst, pkt));
139
140 delete state;
141}
142
143template <class Impl>
144LSQUnit<Impl>::LSQUnit()
145 : loads(0), stores(0), storesToWB(0), cacheBlockMask(0), stalled(false),
146 isStoreBlocked(false), storeInFlight(false), hasPendingPkt(false),
147 pendingPkt(nullptr)
148{
149}
150
151template<class Impl>
152void
153LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params,
154 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries,
155 unsigned id)
156{
157 cpu = cpu_ptr;
158 iewStage = iew_ptr;
159
160 lsq = lsq_ptr;
161
162 lsqID = id;
163
164 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id);
165
166 // Add 1 for the sentinel entry (they are circular queues).
167 LQEntries = maxLQEntries + 1;
168 SQEntries = maxSQEntries + 1;
169
170 //Due to uint8_t index in LSQSenderState
171 assert(LQEntries <= 256);
172 assert(SQEntries <= 256);
173
174 loadQueue.resize(LQEntries);
175 storeQueue.resize(SQEntries);
176
177 depCheckShift = params->LSQDepCheckShift;
178 checkLoads = params->LSQCheckLoads;
179 cachePorts = params->cachePorts;
180 needsTSO = params->needsTSO;
181
182 resetState();
183}
184
185
186template<class Impl>
187void
188LSQUnit<Impl>::resetState()
189{
190 loads = stores = storesToWB = 0;
191
192 loadHead = loadTail = 0;
193
194 storeHead = storeWBIdx = storeTail = 0;
195
196 usedPorts = 0;
197
198 retryPkt = NULL;
199 memDepViolator = NULL;
200
201 stalled = false;
202
203 cacheBlockMask = ~(cpu->cacheLineSize() - 1);
204}
205
206template<class Impl>
207std::string
208LSQUnit<Impl>::name() const
209{
210 if (Impl::MaxThreads == 1) {
211 return iewStage->name() + ".lsq";
212 } else {
213 return iewStage->name() + ".lsq.thread" + std::to_string(lsqID);
214 }
215}
216
217template<class Impl>
218void
219LSQUnit<Impl>::regStats()
220{
221 lsqForwLoads
222 .name(name() + ".forwLoads")
223 .desc("Number of loads that had data forwarded from stores");
224
225 invAddrLoads
226 .name(name() + ".invAddrLoads")
227 .desc("Number of loads ignored due to an invalid address");
228
229 lsqSquashedLoads
230 .name(name() + ".squashedLoads")
231 .desc("Number of loads squashed");
232
233 lsqIgnoredResponses
234 .name(name() + ".ignoredResponses")
235 .desc("Number of memory responses ignored because the instruction is squashed");
236
237 lsqMemOrderViolation
238 .name(name() + ".memOrderViolation")
239 .desc("Number of memory ordering violations");
240
241 lsqSquashedStores
242 .name(name() + ".squashedStores")
243 .desc("Number of stores squashed");
244
245 invAddrSwpfs
246 .name(name() + ".invAddrSwpfs")
247 .desc("Number of software prefetches ignored due to an invalid address");
248
249 lsqBlockedLoads
250 .name(name() + ".blockedLoads")
251 .desc("Number of blocked loads due to partial load-store forwarding");
252
253 lsqRescheduledLoads
254 .name(name() + ".rescheduledLoads")
255 .desc("Number of loads that were rescheduled");
256
257 lsqCacheBlocked
258 .name(name() + ".cacheBlocked")
259 .desc("Number of times an access to memory failed due to the cache being blocked");
260}
261
262template<class Impl>
263void
264LSQUnit<Impl>::setDcachePort(MasterPort *dcache_port)
265{
266 dcachePort = dcache_port;
267}
268
269template<class Impl>
270void
271LSQUnit<Impl>::clearLQ()
272{
273 loadQueue.clear();
274}
275
276template<class Impl>
277void
278LSQUnit<Impl>::clearSQ()
279{
280 storeQueue.clear();
281}
282
283template<class Impl>
284void
285LSQUnit<Impl>::drainSanityCheck() const
286{
287 for (int i = 0; i < loadQueue.size(); ++i)
288 assert(!loadQueue[i]);
289
290 assert(storesToWB == 0);
291 assert(!retryPkt);
292}
293
294template<class Impl>
295void
296LSQUnit<Impl>::takeOverFrom()
297{
298 resetState();
299}
300
301template<class Impl>
302void
303LSQUnit<Impl>::resizeLQ(unsigned size)
304{
305 unsigned size_plus_sentinel = size + 1;
306 assert(size_plus_sentinel >= LQEntries);
307
308 if (size_plus_sentinel > LQEntries) {
309 while (size_plus_sentinel > loadQueue.size()) {
310 DynInstPtr dummy;
311 loadQueue.push_back(dummy);
312 LQEntries++;
313 }
314 } else {
315 LQEntries = size_plus_sentinel;
316 }
317
318 assert(LQEntries <= 256);
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 assert(SQEntries <= 256);
337}
338
339template <class Impl>
340void
341LSQUnit<Impl>::insert(DynInstPtr &inst)
342{
343 assert(inst->isMemRef());
344
345 assert(inst->isLoad() || inst->isStore());
346
347 if (inst->isLoad()) {
348 insertLoad(inst);
349 } else {
350 insertStore(inst);
351 }
352
353 inst->setInLSQ();
354}
355
356template <class Impl>
357void
358LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst)
359{
360 assert((loadTail + 1) % LQEntries != loadHead);
361 assert(loads < LQEntries);
362
363 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n",
364 load_inst->pcState(), loadTail, load_inst->seqNum);
365
366 load_inst->lqIdx = loadTail;
367
368 if (stores == 0) {
369 load_inst->sqIdx = -1;
370 } else {
371 load_inst->sqIdx = storeTail;
372 }
373
374 loadQueue[loadTail] = load_inst;
375
376 incrLdIdx(loadTail);
377
378 ++loads;
379}
380
381template <class Impl>
382void
383LSQUnit<Impl>::insertStore(DynInstPtr &store_inst)
384{
385 // Make sure it is not full before inserting an instruction.
386 assert((storeTail + 1) % SQEntries != storeHead);
387 assert(stores < SQEntries);
388
389 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n",
390 store_inst->pcState(), storeTail, store_inst->seqNum);
391
392 store_inst->sqIdx = storeTail;
393 store_inst->lqIdx = loadTail;
394
395 storeQueue[storeTail] = SQEntry(store_inst);
396
397 incrStIdx(storeTail);
398
399 ++stores;
400}
401
402template <class Impl>
403typename Impl::DynInstPtr
404LSQUnit<Impl>::getMemDepViolator()
405{
406 DynInstPtr temp = memDepViolator;
407
408 memDepViolator = NULL;
409
410 return temp;
411}
412
413template <class Impl>
414unsigned
415LSQUnit<Impl>::numFreeLoadEntries()
416{
417 //LQ has an extra dummy entry to differentiate
418 //empty/full conditions. Subtract 1 from the free entries.
419 DPRINTF(LSQUnit, "LQ size: %d, #loads occupied: %d\n", LQEntries, loads);
420 return LQEntries - loads - 1;
421}
422
423template <class Impl>
424unsigned
425LSQUnit<Impl>::numFreeStoreEntries()
426{
427 //SQ has an extra dummy entry to differentiate
428 //empty/full conditions. Subtract 1 from the free entries.
429 DPRINTF(LSQUnit, "SQ size: %d, #stores occupied: %d\n", SQEntries, stores);
430 return SQEntries - stores - 1;
431
432 }
433
434template <class Impl>
435void
436LSQUnit<Impl>::checkSnoop(PacketPtr pkt)
437{
438 // Should only ever get invalidations in here
439 assert(pkt->isInvalidate());
440
441 int load_idx = loadHead;
442 DPRINTF(LSQUnit, "Got snoop for address %#x\n", pkt->getAddr());
443
444 // Only Invalidate packet calls checkSnoop
445 assert(pkt->isInvalidate());
446 for (int x = 0; x < cpu->numContexts(); x++) {
447 ThreadContext *tc = cpu->getContext(x);
448 bool no_squash = cpu->thread[x]->noSquashFromTC;
449 cpu->thread[x]->noSquashFromTC = true;
450 TheISA::handleLockedSnoop(tc, pkt, cacheBlockMask);
451 cpu->thread[x]->noSquashFromTC = no_squash;
452 }
453
454 Addr invalidate_addr = pkt->getAddr() & cacheBlockMask;
455
456 DynInstPtr ld_inst = loadQueue[load_idx];
457 if (ld_inst) {
458 Addr load_addr_low = ld_inst->physEffAddrLow & cacheBlockMask;
459 Addr load_addr_high = ld_inst->physEffAddrHigh & cacheBlockMask;
460
461 // Check that this snoop didn't just invalidate our lock flag
462 if (ld_inst->effAddrValid() && (load_addr_low == invalidate_addr
463 || load_addr_high == invalidate_addr)
464 && ld_inst->memReqFlags & Request::LLSC)
465 TheISA::handleLockedSnoopHit(ld_inst.get());
466 }
467
468 // If this is the only load in the LSQ we don't care
469 if (load_idx == loadTail)
470 return;
471
472 incrLdIdx(load_idx);
473
474 bool force_squash = false;
475
476 while (load_idx != loadTail) {
477 DynInstPtr ld_inst = loadQueue[load_idx];
478
479 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) {
480 incrLdIdx(load_idx);
481 continue;
482 }
483
484 Addr load_addr_low = ld_inst->physEffAddrLow & cacheBlockMask;
485 Addr load_addr_high = ld_inst->physEffAddrHigh & cacheBlockMask;
486
487 DPRINTF(LSQUnit, "-- inst [sn:%lli] load_addr: %#x to pktAddr:%#x\n",
488 ld_inst->seqNum, load_addr_low, invalidate_addr);
489
490 if ((load_addr_low == invalidate_addr
491 || load_addr_high == invalidate_addr) || force_squash) {
492 if (needsTSO) {
493 // If we have a TSO system, as all loads must be ordered with
494 // all other loads, this load as well as *all* subsequent loads
495 // need to be squashed to prevent possible load reordering.
496 force_squash = true;
497 }
498 if (ld_inst->possibleLoadViolation() || force_squash) {
499 DPRINTF(LSQUnit, "Conflicting load at addr %#x [sn:%lli]\n",
500 pkt->getAddr(), ld_inst->seqNum);
501
502 // Mark the load for re-execution
503 ld_inst->fault = std::make_shared<ReExec>();
504 } else {
505 DPRINTF(LSQUnit, "HitExternal Snoop for addr %#x [sn:%lli]\n",
506 pkt->getAddr(), ld_inst->seqNum);
507
508 // Make sure that we don't lose a snoop hitting a LOCKED
509 // address since the LOCK* flags don't get updated until
510 // commit.
511 if (ld_inst->memReqFlags & Request::LLSC)
512 TheISA::handleLockedSnoopHit(ld_inst.get());
513
514 // If a older load checks this and it's true
515 // then we might have missed the snoop
516 // in which case we need to invalidate to be sure
517 ld_inst->hitExternalSnoop(true);
518 }
519 }
520 incrLdIdx(load_idx);
521 }
522 return;
523}
524
525template <class Impl>
526Fault
527LSQUnit<Impl>::checkViolations(int load_idx, DynInstPtr &inst)
528{
529 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift;
530 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift;
531
532 /** @todo in theory you only need to check an instruction that has executed
533 * however, there isn't a good way in the pipeline at the moment to check
534 * all instructions that will execute before the store writes back. Thus,
535 * like the implementation that came before it, we're overly conservative.
536 */
537 while (load_idx != loadTail) {
538 DynInstPtr ld_inst = loadQueue[load_idx];
539 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) {
540 incrLdIdx(load_idx);
541 continue;
542 }
543
544 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift;
545 Addr ld_eff_addr2 =
546 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift;
547
548 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) {
549 if (inst->isLoad()) {
550 // If this load is to the same block as an external snoop
551 // invalidate that we've observed then the load needs to be
552 // squashed as it could have newer data
553 if (ld_inst->hitExternalSnoop()) {
554 if (!memDepViolator ||
555 ld_inst->seqNum < memDepViolator->seqNum) {
556 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] "
557 "and [sn:%lli] at address %#x\n",
558 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
559 memDepViolator = ld_inst;
560
561 ++lsqMemOrderViolation;
562
563 return std::make_shared<GenericISA::M5PanicFault>(
564 "Detected fault with inst [sn:%lli] and "
565 "[sn:%lli] at address %#x\n",
566 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
567 }
568 }
569
570 // Otherwise, mark the load has a possible load violation
571 // and if we see a snoop before it's commited, we need to squash
572 ld_inst->possibleLoadViolation(true);
573 DPRINTF(LSQUnit, "Found possible load violation at addr: %#x"
574 " between instructions [sn:%lli] and [sn:%lli]\n",
575 inst_eff_addr1, inst->seqNum, ld_inst->seqNum);
576 } else {
577 // A load/store incorrectly passed this store.
578 // Check if we already have a violator, or if it's newer
579 // squash and refetch.
580 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum)
581 break;
582
583 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and "
584 "[sn:%lli] at address %#x\n",
585 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
586 memDepViolator = ld_inst;
587
588 ++lsqMemOrderViolation;
589
590 return std::make_shared<GenericISA::M5PanicFault>(
591 "Detected fault with "
592 "inst [sn:%lli] and [sn:%lli] at address %#x\n",
593 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
594 }
595 }
596
597 incrLdIdx(load_idx);
598 }
599 return NoFault;
600}
601
602
603
604
605template <class Impl>
606Fault
607LSQUnit<Impl>::executeLoad(DynInstPtr &inst)
608{
609 using namespace TheISA;
610 // Execute a specific load.
611 Fault load_fault = NoFault;
612
613 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n",
614 inst->pcState(), inst->seqNum);
615
616 assert(!inst->isSquashed());
617
618 load_fault = inst->initiateAcc();
619
620 if (inst->isTranslationDelayed() &&
621 load_fault == NoFault)
622 return load_fault;
623
624 // If the instruction faulted or predicated false, then we need to send it
625 // along to commit without the instruction completing.
626 if (load_fault != NoFault || !inst->readPredicate()) {
627 // Send this instruction to commit, also make sure iew stage
628 // realizes there is activity. Mark it as executed unless it
629 // is a strictly ordered load that needs to hit the head of
630 // commit.
631 if (!inst->readPredicate())
632 inst->forwardOldRegs();
633 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n",
634 inst->seqNum,
635 (load_fault != NoFault ? "fault" : "predication"));
636 if (!(inst->hasRequest() && inst->strictlyOrdered()) ||
637 inst->isAtCommit()) {
638 inst->setExecuted();
639 }
640 iewStage->instToCommit(inst);
641 iewStage->activityThisCycle();
642 } else {
643 assert(inst->effAddrValid());
644 int load_idx = inst->lqIdx;
645 incrLdIdx(load_idx);
646
647 if (checkLoads)
648 return checkViolations(load_idx, inst);
649 }
650
651 return load_fault;
652}
653
654template <class Impl>
655Fault
656LSQUnit<Impl>::executeStore(DynInstPtr &store_inst)
657{
658 using namespace TheISA;
659 // Make sure that a store exists.
660 assert(stores != 0);
661
662 int store_idx = store_inst->sqIdx;
663
664 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n",
665 store_inst->pcState(), store_inst->seqNum);
666
667 assert(!store_inst->isSquashed());
668
669 // Check the recently completed loads to see if any match this store's
670 // address. If so, then we have a memory ordering violation.
671 int load_idx = store_inst->lqIdx;
672
673 Fault store_fault = store_inst->initiateAcc();
674
675 if (store_inst->isTranslationDelayed() &&
676 store_fault == NoFault)
677 return store_fault;
678
679 if (!store_inst->readPredicate())
680 store_inst->forwardOldRegs();
681
682 if (storeQueue[store_idx].size == 0) {
683 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n",
684 store_inst->pcState(), store_inst->seqNum);
685
686 return store_fault;
687 } else if (!store_inst->readPredicate()) {
688 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n",
689 store_inst->seqNum);
690 return store_fault;
691 }
692
693 assert(store_fault == NoFault);
694
695 if (store_inst->isStoreConditional()) {
696 // Store conditionals need to set themselves as able to
697 // writeback if we haven't had a fault by here.
698 storeQueue[store_idx].canWB = true;
699
700 ++storesToWB;
701 }
702
703 return checkViolations(load_idx, store_inst);
704
705}
706
707template <class Impl>
708void
709LSQUnit<Impl>::commitLoad()
710{
711 assert(loadQueue[loadHead]);
712
713 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n",
714 loadQueue[loadHead]->pcState());
715
716 loadQueue[loadHead] = NULL;
717
718 incrLdIdx(loadHead);
719
720 --loads;
721}
722
723template <class Impl>
724void
725LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst)
726{
727 assert(loads == 0 || loadQueue[loadHead]);
728
729 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) {
730 commitLoad();
731 }
732}
733
734template <class Impl>
735void
736LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst)
737{
738 assert(stores == 0 || storeQueue[storeHead].inst);
739
740 int store_idx = storeHead;
741
742 while (store_idx != storeTail) {
743 assert(storeQueue[store_idx].inst);
744 // Mark any stores that are now committed and have not yet
745 // been marked as able to write back.
746 if (!storeQueue[store_idx].canWB) {
747 if (storeQueue[store_idx].inst->seqNum > youngest_inst) {
748 break;
749 }
750 DPRINTF(LSQUnit, "Marking store as able to write back, PC "
751 "%s [sn:%lli]\n",
752 storeQueue[store_idx].inst->pcState(),
753 storeQueue[store_idx].inst->seqNum);
754
755 storeQueue[store_idx].canWB = true;
756
757 ++storesToWB;
758 }
759
760 incrStIdx(store_idx);
761 }
762}
763
764template <class Impl>
765void
766LSQUnit<Impl>::writebackPendingStore()
767{
768 if (hasPendingPkt) {
769 assert(pendingPkt != NULL);
770
771 // If the cache is blocked, this will store the packet for retry.
772 if (sendStore(pendingPkt)) {
773 storePostSend(pendingPkt);
774 }
775 pendingPkt = NULL;
776 hasPendingPkt = false;
777 }
778}
779
780template <class Impl>
781void
782LSQUnit<Impl>::writebackStores()
783{
784 // First writeback the second packet from any split store that didn't
785 // complete last cycle because there weren't enough cache ports available.
786 if (TheISA::HasUnalignedMemAcc) {
787 writebackPendingStore();
788 }
789
790 while (storesToWB > 0 &&
791 storeWBIdx != storeTail &&
792 storeQueue[storeWBIdx].inst &&
793 storeQueue[storeWBIdx].canWB &&
794 ((!needsTSO) || (!storeInFlight)) &&
795 usedPorts < cachePorts) {
796
797 if (isStoreBlocked) {
798 DPRINTF(LSQUnit, "Unable to write back any more stores, cache"
799 " is blocked!\n");
800 break;
801 }
802
803 // Store didn't write any data so no need to write it back to
804 // memory.
805 if (storeQueue[storeWBIdx].size == 0) {
806 completeStore(storeWBIdx);
807
808 incrStIdx(storeWBIdx);
809
810 continue;
811 }
812
813 ++usedPorts;
814
815 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) {
816 incrStIdx(storeWBIdx);
817
818 continue;
819 }
820
821 assert(storeQueue[storeWBIdx].req);
822 assert(!storeQueue[storeWBIdx].committed);
823
824 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) {
825 assert(storeQueue[storeWBIdx].sreqLow);
826 assert(storeQueue[storeWBIdx].sreqHigh);
827 }
828
829 DynInstPtr inst = storeQueue[storeWBIdx].inst;
830
831 Request *req = storeQueue[storeWBIdx].req;
832 RequestPtr sreqLow = storeQueue[storeWBIdx].sreqLow;
833 RequestPtr sreqHigh = storeQueue[storeWBIdx].sreqHigh;
834
835 storeQueue[storeWBIdx].committed = true;
836
837 assert(!inst->memData);
838 inst->memData = new uint8_t[req->getSize()];
839
840 if (storeQueue[storeWBIdx].isAllZeros)
841 memset(inst->memData, 0, req->getSize());
842 else
843 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize());
844
845 PacketPtr data_pkt;
846 PacketPtr snd_data_pkt = NULL;
847
848 LSQSenderState *state = new LSQSenderState;
849 state->isLoad = false;
850 state->idx = storeWBIdx;
851 state->inst = inst;
852
853 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) {
854
855 // Build a single data packet if the store isn't split.
856 data_pkt = Packet::createWrite(req);
857 data_pkt->dataStatic(inst->memData);
858 data_pkt->senderState = state;
859 } else {
860 // Create two packets if the store is split in two.
861 data_pkt = Packet::createWrite(sreqLow);
862 snd_data_pkt = Packet::createWrite(sreqHigh);
863
864 data_pkt->dataStatic(inst->memData);
865 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize());
866
867 data_pkt->senderState = state;
868 snd_data_pkt->senderState = state;
869
870 state->isSplit = true;
871 state->outstanding = 2;
872
873 // Can delete the main request now.
874 delete req;
875 req = sreqLow;
876 }
877
878 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s "
879 "to Addr:%#x, data:%#x [sn:%lli]\n",
880 storeWBIdx, inst->pcState(),
881 req->getPaddr(), (int)*(inst->memData),
882 inst->seqNum);
883
884 // @todo: Remove this SC hack once the memory system handles it.
885 if (inst->isStoreConditional()) {
886 assert(!storeQueue[storeWBIdx].isSplit);
887 // Disable recording the result temporarily. Writing to
888 // misc regs normally updates the result, but this is not
889 // the desired behavior when handling store conditionals.
890 inst->recordResult(false);
891 bool success = TheISA::handleLockedWrite(inst.get(), req, cacheBlockMask);
892 inst->recordResult(true);
893
894 if (!success) {
895 // Instantly complete this store.
896 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. "
897 "Instantly completing it.\n",
898 inst->seqNum);
899 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this);
900 cpu->schedule(wb, curTick() + 1);
901 if (cpu->checker) {
902 // Make sure to set the LLSC data for verification
903 // if checker is loaded
904 inst->reqToVerify->setExtraData(0);
905 inst->completeAcc(data_pkt);
906 }
907 completeStore(storeWBIdx);
908 incrStIdx(storeWBIdx);
909 continue;
910 }
911 } else {
912 // Non-store conditionals do not need a writeback.
913 state->noWB = true;
914 }
915
916 bool split =
917 TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit;
918
919 ThreadContext *thread = cpu->tcBase(lsqID);
920
921 if (req->isMmappedIpr()) {
922 assert(!inst->isStoreConditional());
923 TheISA::handleIprWrite(thread, data_pkt);
924 delete data_pkt;
925 if (split) {
926 assert(snd_data_pkt->req->isMmappedIpr());
927 TheISA::handleIprWrite(thread, snd_data_pkt);
928 delete snd_data_pkt;
929 delete sreqLow;
930 delete sreqHigh;
931 }
932 delete state;
933 delete req;
934 completeStore(storeWBIdx);
935 incrStIdx(storeWBIdx);
936 } else if (!sendStore(data_pkt)) {
937 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will"
938 "retry later\n",
939 inst->seqNum);
940
941 // Need to store the second packet, if split.
942 if (split) {
943 state->pktToSend = true;
944 state->pendingPacket = snd_data_pkt;
945 }
946 } else {
947
948 // If split, try to send the second packet too
949 if (split) {
950 assert(snd_data_pkt);
951
952 // Ensure there are enough ports to use.
953 if (usedPorts < cachePorts) {
954 ++usedPorts;
955 if (sendStore(snd_data_pkt)) {
956 storePostSend(snd_data_pkt);
957 } else {
958 DPRINTF(IEW, "D-Cache became blocked when writing"
959 " [sn:%lli] second packet, will retry later\n",
960 inst->seqNum);
961 }
962 } else {
963
964 // Store the packet for when there's free ports.
965 assert(pendingPkt == NULL);
966 pendingPkt = snd_data_pkt;
967 hasPendingPkt = true;
968 }
969 } else {
970
971 // Not a split store.
972 storePostSend(data_pkt);
973 }
974 }
975 }
976
977 // Not sure this should set it to 0.
978 usedPorts = 0;
979
980 assert(stores >= 0 && storesToWB >= 0);
981}
982
983/*template <class Impl>
984void
985LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum)
986{
987 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(),
988 mshrSeqNums.end(),
989 seqNum);
990
991 if (mshr_it != mshrSeqNums.end()) {
992 mshrSeqNums.erase(mshr_it);
993 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size());
994 }
995}*/
996
997template <class Impl>
998void
999LSQUnit<Impl>::squash(const InstSeqNum &squashed_num)
1000{
1001 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!"
1002 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores);
1003
1004 int load_idx = loadTail;
1005 decrLdIdx(load_idx);
1006
1007 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) {
1008 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, "
1009 "[sn:%lli]\n",
1010 loadQueue[load_idx]->pcState(),
1011 loadQueue[load_idx]->seqNum);
1012
1013 if (isStalled() && load_idx == stallingLoadIdx) {
1014 stalled = false;
1015 stallingStoreIsn = 0;
1016 stallingLoadIdx = 0;
1017 }
1018
1019 // Clear the smart pointer to make sure it is decremented.
1020 loadQueue[load_idx]->setSquashed();
1021 loadQueue[load_idx] = NULL;
1022 --loads;
1023
1024 // Inefficient!
1025 loadTail = load_idx;
1026
1027 decrLdIdx(load_idx);
1028 ++lsqSquashedLoads;
1029 }
1030
1031 if (memDepViolator && squashed_num < memDepViolator->seqNum) {
1032 memDepViolator = NULL;
1033 }
1034
1035 int store_idx = storeTail;
1036 decrStIdx(store_idx);
1037
1038 while (stores != 0 &&
1039 storeQueue[store_idx].inst->seqNum > squashed_num) {
1040 // Instructions marked as can WB are already committed.
1041 if (storeQueue[store_idx].canWB) {
1042 break;
1043 }
1044
1045 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, "
1046 "idx:%i [sn:%lli]\n",
1047 storeQueue[store_idx].inst->pcState(),
1048 store_idx, storeQueue[store_idx].inst->seqNum);
1049
1050 // I don't think this can happen. It should have been cleared
1051 // by the stalling load.
1052 if (isStalled() &&
1053 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) {
1054 panic("Is stalled should have been cleared by stalling load!\n");
1055 stalled = false;
1056 stallingStoreIsn = 0;
1057 }
1058
1059 // Clear the smart pointer to make sure it is decremented.
1060 storeQueue[store_idx].inst->setSquashed();
1061 storeQueue[store_idx].inst = NULL;
1062 storeQueue[store_idx].canWB = 0;
1063
1064 // Must delete request now that it wasn't handed off to
1065 // memory. This is quite ugly. @todo: Figure out the proper
1066 // place to really handle request deletes.
1067 delete storeQueue[store_idx].req;
1068 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) {
1069 delete storeQueue[store_idx].sreqLow;
1070 delete storeQueue[store_idx].sreqHigh;
1071
1072 storeQueue[store_idx].sreqLow = NULL;
1073 storeQueue[store_idx].sreqHigh = NULL;
1074 }
1075
1076 storeQueue[store_idx].req = NULL;
1077 --stores;
1078
1079 // Inefficient!
1080 storeTail = store_idx;
1081
1082 decrStIdx(store_idx);
1083 ++lsqSquashedStores;
1084 }
1085}
1086
1087template <class Impl>
1088void
1089LSQUnit<Impl>::storePostSend(PacketPtr pkt)
1090{
1091 if (isStalled() &&
1092 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) {
1093 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] "
1094 "load idx:%i\n",
1095 stallingStoreIsn, stallingLoadIdx);
1096 stalled = false;
1097 stallingStoreIsn = 0;
1098 iewStage->replayMemInst(loadQueue[stallingLoadIdx]);
1099 }
1100
1101 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) {
1102 // The store is basically completed at this time. This
1103 // only works so long as the checker doesn't try to
1104 // verify the value in memory for stores.
1105 storeQueue[storeWBIdx].inst->setCompleted();
1106
1107 if (cpu->checker) {
1108 cpu->checker->verify(storeQueue[storeWBIdx].inst);
1109 }
1110 }
1111
1112 if (needsTSO) {
1113 storeInFlight = true;
1114 }
1115
1116 incrStIdx(storeWBIdx);
1117}
1118
1119template <class Impl>
1120void
1121LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt)
1122{
1123 iewStage->wakeCPU();
1124
1125 // Squashed instructions do not need to complete their access.
1126 if (inst->isSquashed()) {
1127 assert(!inst->isStore());
1128 ++lsqIgnoredResponses;
1129 return;
1130 }
1131
1132 if (!inst->isExecuted()) {
1133 inst->setExecuted();
1134
1135 if (inst->fault == NoFault) {
1136 // Complete access to copy data to proper place.
1137 inst->completeAcc(pkt);
1138 } else {
1139 // If the instruction has an outstanding fault, we cannot complete
1140 // the access as this discards the current fault.
1141
1142 // If we have an outstanding fault, the fault should only be of
1143 // type ReExec.
1144 assert(dynamic_cast<ReExec*>(inst->fault.get()) != nullptr);
1145
1146 DPRINTF(LSQUnit, "Not completing instruction [sn:%lli] access "
1147 "due to pending fault.\n", inst->seqNum);
1148 }
1149 }
1150
1151 // Need to insert instruction into queue to commit
1152 iewStage->instToCommit(inst);
1153
1154 iewStage->activityThisCycle();
1155
1156 // see if this load changed the PC
1157 iewStage->checkMisprediction(inst);
1158}
1159
1160template <class Impl>
1161void
1162LSQUnit<Impl>::completeStore(int store_idx)
1163{
1164 assert(storeQueue[store_idx].inst);
1165 storeQueue[store_idx].completed = true;
1166 --storesToWB;
1167 // A bit conservative because a store completion may not free up entries,
1168 // but hopefully avoids two store completions in one cycle from making
1169 // the CPU tick twice.
1170 cpu->wakeCPU();
1171 cpu->activityThisCycle();
1172
1173 if (store_idx == storeHead) {
1174 do {
1175 incrStIdx(storeHead);
1176
1177 --stores;
1178 } while (storeQueue[storeHead].completed &&
1179 storeHead != storeTail);
1180
1181 iewStage->updateLSQNextCycle = true;
1182 }
1183
1184 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head "
1185 "idx:%i\n",
1186 storeQueue[store_idx].inst->seqNum, store_idx, storeHead);
1187
1188#if TRACING_ON
1189 if (DTRACE(O3PipeView)) {
1190 storeQueue[store_idx].inst->storeTick =
1191 curTick() - storeQueue[store_idx].inst->fetchTick;
1192 }
1193#endif
1194
1195 if (isStalled() &&
1196 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) {
1197 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] "
1198 "load idx:%i\n",
1199 stallingStoreIsn, stallingLoadIdx);
1200 stalled = false;
1201 stallingStoreIsn = 0;
1202 iewStage->replayMemInst(loadQueue[stallingLoadIdx]);
1203 }
1204
1205 storeQueue[store_idx].inst->setCompleted();
1206
1207 if (needsTSO) {
1208 storeInFlight = false;
1209 }
1210
1211 // Tell the checker we've completed this instruction. Some stores
1212 // may get reported twice to the checker, but the checker can
1213 // handle that case.
1214 if (cpu->checker) {
1215 cpu->checker->verify(storeQueue[store_idx].inst);
1216 }
1217}
1218
1219template <class Impl>
1220bool
1221LSQUnit<Impl>::sendStore(PacketPtr data_pkt)
1222{
1223 if (!dcachePort->sendTimingReq(data_pkt)) {
1224 // Need to handle becoming blocked on a store.
1225 isStoreBlocked = true;
1226 ++lsqCacheBlocked;
1227 assert(retryPkt == NULL);
1228 retryPkt = data_pkt;
1229 return false;
1230 }
1231 return true;
1232}
1233
1234template <class Impl>
1235void
1236LSQUnit<Impl>::recvRetry()
1237{
1238 if (isStoreBlocked) {
1239 DPRINTF(LSQUnit, "Receiving retry: store blocked\n");
1240 assert(retryPkt != NULL);
1241
1242 LSQSenderState *state =
1243 dynamic_cast<LSQSenderState *>(retryPkt->senderState);
1244
1245 if (dcachePort->sendTimingReq(retryPkt)) {
1246 // Don't finish the store unless this is the last packet.
1247 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend ||
1248 state->pendingPacket == retryPkt) {
1249 state->pktToSend = false;
1250 storePostSend(retryPkt);
1251 }
1252 retryPkt = NULL;
1253 isStoreBlocked = false;
1254
1255 // Send any outstanding packet.
1256 if (TheISA::HasUnalignedMemAcc && state->pktToSend) {
1257 assert(state->pendingPacket);
1258 if (sendStore(state->pendingPacket)) {
1259 storePostSend(state->pendingPacket);
1260 }
1261 }
1262 } else {
1263 // Still blocked!
1264 ++lsqCacheBlocked;
1265 }
1266 }
1267}
1268
1269template <class Impl>
1270inline void
1271LSQUnit<Impl>::incrStIdx(int &store_idx) const
1272{
1273 if (++store_idx >= SQEntries)
1274 store_idx = 0;
1275}
1276
1277template <class Impl>
1278inline void
1279LSQUnit<Impl>::decrStIdx(int &store_idx) const
1280{
1281 if (--store_idx < 0)
1282 store_idx += SQEntries;
1283}
1284
1285template <class Impl>
1286inline void
1287LSQUnit<Impl>::incrLdIdx(int &load_idx) const
1288{
1289 if (++load_idx >= LQEntries)
1290 load_idx = 0;
1291}
1292
1293template <class Impl>
1294inline void
1295LSQUnit<Impl>::decrLdIdx(int &load_idx) const
1296{
1297 if (--load_idx < 0)
1298 load_idx += LQEntries;
1299}
1300
1301template <class Impl>
1302void
1303LSQUnit<Impl>::dumpInsts() const
1304{
1305 cprintf("Load store queue: Dumping instructions.\n");
1306 cprintf("Load queue size: %i\n", loads);
1307 cprintf("Load queue: ");
1308
1309 int load_idx = loadHead;
1310
1311 while (load_idx != loadTail && loadQueue[load_idx]) {
1312 const DynInstPtr &inst(loadQueue[load_idx]);
1313 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum);
1314
1315 incrLdIdx(load_idx);
1316 }
1317 cprintf("\n");
1318
1319 cprintf("Store queue size: %i\n", stores);
1320 cprintf("Store queue: ");
1321
1322 int store_idx = storeHead;
1323
1324 while (store_idx != storeTail && storeQueue[store_idx].inst) {
1325 const DynInstPtr &inst(storeQueue[store_idx].inst);
1326 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum);
1327
1328 incrStIdx(store_idx);
1329 }
1330
1331 cprintf("\n");
1332}
1333
1334#endif//__CPU_O3_LSQ_UNIT_IMPL_HH__