1
2/*
3 * Copyright (c) 2010-2013 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 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum);
103
104 // If this is a split access, wait until all packets are received.
105 if (TheISA::HasUnalignedMemAcc && !state->complete()) {
106 delete pkt->req;
107 delete pkt;
108 return;
109 }
110
111 assert(!cpu->switchedOut());
112 if (inst->isSquashed()) {
113 iewStage->decrWb(inst->seqNum);
114 } else {
115 if (!state->noWB) {
116 if (!TheISA::HasUnalignedMemAcc || !state->isSplit ||
117 !state->isLoad) {
118 writeback(inst, pkt);
119 } else {
120 writeback(inst, state->mainPkt);
121 }
122 }
123
124 if (inst->isStore()) {
125 completeStore(state->idx);
126 }
127 }
128
129 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) {
130 delete state->mainPkt->req;
131 delete state->mainPkt;
132 }
133
134 pkt->req->setAccessLatency();
135 cpu->ppDataAccessComplete->notify(std::make_pair(inst, pkt));
136
137 delete state;
138 delete pkt->req;
139 delete pkt;
140}
141
142template <class Impl>
143LSQUnit<Impl>::LSQUnit()
144 : loads(0), stores(0), storesToWB(0), cacheBlockMask(0), stalled(false),
145 isStoreBlocked(false), isLoadBlocked(false),
146 loadBlockedHandled(false), storeInFlight(false), hasPendingPkt(false)
147{
148}
149
150template<class Impl>
151void
152LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params,
153 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries,
154 unsigned id)
155{
156 cpu = cpu_ptr;
157 iewStage = iew_ptr;
158
159 lsq = lsq_ptr;
160
161 lsqID = id;
162
163 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id);
164
165 // Add 1 for the sentinel entry (they are circular queues).
166 LQEntries = maxLQEntries + 1;
167 SQEntries = maxSQEntries + 1;
168
169 //Due to uint8_t index in LSQSenderState
170 assert(LQEntries <= 256);
171 assert(SQEntries <= 256);
172
173 loadQueue.resize(LQEntries);
174 storeQueue.resize(SQEntries);
175
176 depCheckShift = params->LSQDepCheckShift;
177 checkLoads = params->LSQCheckLoads;
178 cachePorts = params->cachePorts;
179 needsTSO = params->needsTSO;
180
181 resetState();
182}
183
184
185template<class Impl>
186void
187LSQUnit<Impl>::resetState()
188{
189 loads = stores = storesToWB = 0;
190
191 loadHead = loadTail = 0;
192
193 storeHead = storeWBIdx = storeTail = 0;
194
195 usedPorts = 0;
196
197 retryPkt = NULL;
198 memDepViolator = NULL;
199
200 blockedLoadSeqNum = 0;
201
202 stalled = false;
203 isLoadBlocked = false;
204 loadBlockedHandled = false;
205
206 cacheBlockMask = ~(cpu->cacheLineSize() - 1);
207}
208
209template<class Impl>
210std::string
211LSQUnit<Impl>::name() const
212{
213 if (Impl::MaxThreads == 1) {
214 return iewStage->name() + ".lsq";
215 } else {
216 return iewStage->name() + ".lsq.thread" + to_string(lsqID);
217 }
218}
219
220template<class Impl>
221void
222LSQUnit<Impl>::regStats()
223{
224 lsqForwLoads
225 .name(name() + ".forwLoads")
226 .desc("Number of loads that had data forwarded from stores");
227
228 invAddrLoads
229 .name(name() + ".invAddrLoads")
230 .desc("Number of loads ignored due to an invalid address");
231
232 lsqSquashedLoads
233 .name(name() + ".squashedLoads")
234 .desc("Number of loads squashed");
235
236 lsqIgnoredResponses
237 .name(name() + ".ignoredResponses")
238 .desc("Number of memory responses ignored because the instruction is squashed");
239
240 lsqMemOrderViolation
241 .name(name() + ".memOrderViolation")
242 .desc("Number of memory ordering violations");
243
244 lsqSquashedStores
245 .name(name() + ".squashedStores")
246 .desc("Number of stores squashed");
247
248 invAddrSwpfs
249 .name(name() + ".invAddrSwpfs")
250 .desc("Number of software prefetches ignored due to an invalid address");
251
252 lsqBlockedLoads
253 .name(name() + ".blockedLoads")
254 .desc("Number of blocked loads due to partial load-store forwarding");
255
256 lsqRescheduledLoads
257 .name(name() + ".rescheduledLoads")
258 .desc("Number of loads that were rescheduled");
259
260 lsqCacheBlocked
261 .name(name() + ".cacheBlocked")
262 .desc("Number of times an access to memory failed due to the cache being blocked");
263}
264
265template<class Impl>
266void
267LSQUnit<Impl>::setDcachePort(MasterPort *dcache_port)
268{
269 dcachePort = dcache_port;
270}
271
272template<class Impl>
273void
274LSQUnit<Impl>::clearLQ()
275{
276 loadQueue.clear();
277}
278
279template<class Impl>
280void
281LSQUnit<Impl>::clearSQ()
282{
283 storeQueue.clear();
284}
285
286template<class Impl>
287void
288LSQUnit<Impl>::drainSanityCheck() const
289{
290 for (int i = 0; i < loadQueue.size(); ++i)
291 assert(!loadQueue[i]);
292
293 assert(storesToWB == 0);
294 assert(!retryPkt);
295}
296
297template<class Impl>
298void
299LSQUnit<Impl>::takeOverFrom()
300{
301 resetState();
302}
303
304template<class Impl>
305void
306LSQUnit<Impl>::resizeLQ(unsigned size)
307{
308 unsigned size_plus_sentinel = size + 1;
309 assert(size_plus_sentinel >= LQEntries);
310
311 if (size_plus_sentinel > LQEntries) {
312 while (size_plus_sentinel > loadQueue.size()) {
313 DynInstPtr dummy;
314 loadQueue.push_back(dummy);
315 LQEntries++;
316 }
317 } else {
318 LQEntries = size_plus_sentinel;
319 }
320
321 assert(LQEntries <= 256);
322}
323
324template<class Impl>
325void
326LSQUnit<Impl>::resizeSQ(unsigned size)
327{
328 unsigned size_plus_sentinel = size + 1;
329 if (size_plus_sentinel > SQEntries) {
330 while (size_plus_sentinel > storeQueue.size()) {
331 SQEntry dummy;
332 storeQueue.push_back(dummy);
333 SQEntries++;
334 }
335 } else {
336 SQEntries = size_plus_sentinel;
337 }
338
339 assert(SQEntries <= 256);
340}
341
342template <class Impl>
343void
344LSQUnit<Impl>::insert(DynInstPtr &inst)
345{
346 assert(inst->isMemRef());
347
348 assert(inst->isLoad() || inst->isStore());
349
350 if (inst->isLoad()) {
351 insertLoad(inst);
352 } else {
353 insertStore(inst);
354 }
355
356 inst->setInLSQ();
357}
358
359template <class Impl>
360void
361LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst)
362{
363 assert((loadTail + 1) % LQEntries != loadHead);
364 assert(loads < LQEntries);
365
366 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n",
367 load_inst->pcState(), loadTail, load_inst->seqNum);
368
369 load_inst->lqIdx = loadTail;
370
371 if (stores == 0) {
372 load_inst->sqIdx = -1;
373 } else {
374 load_inst->sqIdx = storeTail;
375 }
376
377 loadQueue[loadTail] = load_inst;
378
379 incrLdIdx(loadTail);
380
381 ++loads;
382}
383
384template <class Impl>
385void
386LSQUnit<Impl>::insertStore(DynInstPtr &store_inst)
387{
388 // Make sure it is not full before inserting an instruction.
389 assert((storeTail + 1) % SQEntries != storeHead);
390 assert(stores < SQEntries);
391
392 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n",
393 store_inst->pcState(), storeTail, store_inst->seqNum);
394
395 store_inst->sqIdx = storeTail;
396 store_inst->lqIdx = loadTail;
397
398 storeQueue[storeTail] = SQEntry(store_inst);
399
400 incrStIdx(storeTail);
401
402 ++stores;
403}
404
405template <class Impl>
406typename Impl::DynInstPtr
407LSQUnit<Impl>::getMemDepViolator()
408{
409 DynInstPtr temp = memDepViolator;
410
411 memDepViolator = NULL;
412
413 return temp;
414}
415
416template <class Impl>
417unsigned
418LSQUnit<Impl>::numFreeLoadEntries()
419{
420 //LQ has an extra dummy entry to differentiate
421 //empty/full conditions. Subtract 1 from the free entries.
422 DPRINTF(LSQUnit, "LQ size: %d, #loads occupied: %d\n", LQEntries, loads);
423 return LQEntries - loads - 1;
424}
425
426template <class Impl>
427unsigned
428LSQUnit<Impl>::numFreeStoreEntries()
429{
430 //SQ has an extra dummy entry to differentiate
431 //empty/full conditions. Subtract 1 from the free entries.
432 DPRINTF(LSQUnit, "SQ size: %d, #stores occupied: %d\n", SQEntries, stores);
433 return SQEntries - stores - 1;
434
435 }
436
437template <class Impl>
438void
439LSQUnit<Impl>::checkSnoop(PacketPtr pkt)
440{
441 int load_idx = loadHead;
442 DPRINTF(LSQUnit, "Got snoop for address %#x\n", pkt->getAddr());
443
444 // Unlock the cpu-local monitor when the CPU sees a snoop to a locked
445 // address. The CPU can speculatively execute a LL operation after a pending
446 // SC operation in the pipeline and that can make the cache monitor the CPU
447 // is connected to valid while it really shouldn't be.
448 for (int x = 0; x < cpu->numContexts(); x++) {
449 ThreadContext *tc = cpu->getContext(x);
450 bool no_squash = cpu->thread[x]->noSquashFromTC;
451 cpu->thread[x]->noSquashFromTC = true;
452 TheISA::handleLockedSnoop(tc, pkt, cacheBlockMask);
453 cpu->thread[x]->noSquashFromTC = no_squash;
454 }
455
456 Addr invalidate_addr = pkt->getAddr() & cacheBlockMask;
457
458 DynInstPtr ld_inst = loadQueue[load_idx];
459 if (ld_inst) {
460 Addr load_addr = ld_inst->physEffAddr & cacheBlockMask;
461 // Check that this snoop didn't just invalidate our lock flag
462 if (ld_inst->effAddrValid() && load_addr == invalidate_addr &&
463 ld_inst->memReqFlags & Request::LLSC)
464 TheISA::handleLockedSnoopHit(ld_inst.get());
465 }
466
467 // If this is the only load in the LSQ we don't care
468 if (load_idx == loadTail)
469 return;
470
471 incrLdIdx(load_idx);
472
473 bool force_squash = false;
474
475 while (load_idx != loadTail) {
476 DynInstPtr ld_inst = loadQueue[load_idx];
477
478 if (!ld_inst->effAddrValid() || ld_inst->uncacheable()) {
479 incrLdIdx(load_idx);
480 continue;
481 }
482
483 Addr load_addr = ld_inst->physEffAddr & cacheBlockMask;
484 DPRINTF(LSQUnit, "-- inst [sn:%lli] load_addr: %#x to pktAddr:%#x\n",
485 ld_inst->seqNum, load_addr, invalidate_addr);
486
487 if (load_addr == invalidate_addr || force_squash) {
488 if (needsTSO) {
489 // If we have a TSO system, as all loads must be ordered with
490 // all other loads, this load as well as *all* subsequent loads
491 // need to be squashed to prevent possible load reordering.
492 force_squash = true;
493 }
494 if (ld_inst->possibleLoadViolation() || force_squash) {
495 DPRINTF(LSQUnit, "Conflicting load at addr %#x [sn:%lli]\n",
496 pkt->getAddr(), ld_inst->seqNum);
497
498 // Mark the load for re-execution
499 ld_inst->fault = new ReExec;
500 } else {
501 DPRINTF(LSQUnit, "HitExternal Snoop for addr %#x [sn:%lli]\n",
502 pkt->getAddr(), ld_inst->seqNum);
503
504 // Make sure that we don't lose a snoop hitting a LOCKED
505 // address since the LOCK* flags don't get updated until
506 // commit.
507 if (ld_inst->memReqFlags & Request::LLSC)
508 TheISA::handleLockedSnoopHit(ld_inst.get());
509
510 // If a older load checks this and it's true
511 // then we might have missed the snoop
512 // in which case we need to invalidate to be sure
513 ld_inst->hitExternalSnoop(true);
514 }
515 }
516 incrLdIdx(load_idx);
517 }
518 return;
519}
520
521template <class Impl>
522Fault
523LSQUnit<Impl>::checkViolations(int load_idx, DynInstPtr &inst)
524{
525 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift;
526 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift;
527
528 /** @todo in theory you only need to check an instruction that has executed
529 * however, there isn't a good way in the pipeline at the moment to check
530 * all instructions that will execute before the store writes back. Thus,
531 * like the implementation that came before it, we're overly conservative.
532 */
533 while (load_idx != loadTail) {
534 DynInstPtr ld_inst = loadQueue[load_idx];
535 if (!ld_inst->effAddrValid() || ld_inst->uncacheable()) {
536 incrLdIdx(load_idx);
537 continue;
538 }
539
540 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift;
541 Addr ld_eff_addr2 =
542 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift;
543
544 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) {
545 if (inst->isLoad()) {
546 // If this load is to the same block as an external snoop
547 // invalidate that we've observed then the load needs to be
548 // squashed as it could have newer data
549 if (ld_inst->hitExternalSnoop()) {
550 if (!memDepViolator ||
551 ld_inst->seqNum < memDepViolator->seqNum) {
552 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] "
553 "and [sn:%lli] at address %#x\n",
554 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
555 memDepViolator = ld_inst;
556
557 ++lsqMemOrderViolation;
558
559 return new GenericISA::M5PanicFault(
560 "Detected fault with inst [sn:%lli] and "
561 "[sn:%lli] at address %#x\n",
562 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
563 }
564 }
565
566 // Otherwise, mark the load has a possible load violation
567 // and if we see a snoop before it's commited, we need to squash
568 ld_inst->possibleLoadViolation(true);
569 DPRINTF(LSQUnit, "Found possible load violaiton at addr: %#x"
570 " between instructions [sn:%lli] and [sn:%lli]\n",
571 inst_eff_addr1, inst->seqNum, ld_inst->seqNum);
572 } else {
573 // A load/store incorrectly passed this store.
574 // Check if we already have a violator, or if it's newer
575 // squash and refetch.
576 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum)
577 break;
578
579 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and "
580 "[sn:%lli] at address %#x\n",
581 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
582 memDepViolator = ld_inst;
583
584 ++lsqMemOrderViolation;
585
586 return new GenericISA::M5PanicFault("Detected fault with "
587 "inst [sn:%lli] and [sn:%lli] at address %#x\n",
588 inst->seqNum, ld_inst->seqNum, ld_eff_addr1);
589 }
590 }
591
592 incrLdIdx(load_idx);
593 }
594 return NoFault;
595}
596
597
598
599
600template <class Impl>
601Fault
602LSQUnit<Impl>::executeLoad(DynInstPtr &inst)
603{
604 using namespace TheISA;
605 // Execute a specific load.
606 Fault load_fault = NoFault;
607
608 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n",
609 inst->pcState(), inst->seqNum);
610
611 assert(!inst->isSquashed());
612
613 load_fault = inst->initiateAcc();
614
615 if (inst->isTranslationDelayed() &&
616 load_fault == NoFault)
617 return load_fault;
618
619 // If the instruction faulted or predicated false, then we need to send it
620 // along to commit without the instruction completing.
621 if (load_fault != NoFault || !inst->readPredicate()) {
622 // Send this instruction to commit, also make sure iew stage
623 // realizes there is activity.
624 // Mark it as executed unless it is an uncached load that
625 // needs to hit the head of commit.
626 if (!inst->readPredicate())
627 inst->forwardOldRegs();
628 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n",
629 inst->seqNum,
630 (load_fault != NoFault ? "fault" : "predication"));
631 if (!(inst->hasRequest() && inst->uncacheable()) ||
632 inst->isAtCommit()) {
633 inst->setExecuted();
634 }
635 iewStage->instToCommit(inst);
636 iewStage->activityThisCycle();
637 } else if (!loadBlocked()) {
638 assert(inst->effAddrValid());
639 int load_idx = inst->lqIdx;
640 incrLdIdx(load_idx);
641
642 if (checkLoads)
643 return checkViolations(load_idx, inst);
644 }
645
646 return load_fault;
647}
648
649template <class Impl>
650Fault
651LSQUnit<Impl>::executeStore(DynInstPtr &store_inst)
652{
653 using namespace TheISA;
654 // Make sure that a store exists.
655 assert(stores != 0);
656
657 int store_idx = store_inst->sqIdx;
658
659 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n",
660 store_inst->pcState(), store_inst->seqNum);
661
662 assert(!store_inst->isSquashed());
663
664 // Check the recently completed loads to see if any match this store's
665 // address. If so, then we have a memory ordering violation.
666 int load_idx = store_inst->lqIdx;
667
668 Fault store_fault = store_inst->initiateAcc();
669
670 if (store_inst->isTranslationDelayed() &&
671 store_fault == NoFault)
672 return store_fault;
673
674 if (!store_inst->readPredicate())
675 store_inst->forwardOldRegs();
676
677 if (storeQueue[store_idx].size == 0) {
678 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n",
679 store_inst->pcState(), store_inst->seqNum);
680
681 return store_fault;
682 } else if (!store_inst->readPredicate()) {
683 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n",
684 store_inst->seqNum);
685 return store_fault;
686 }
687
688 assert(store_fault == NoFault);
689
690 if (store_inst->isStoreConditional()) {
691 // Store conditionals need to set themselves as able to
692 // writeback if we haven't had a fault by here.
693 storeQueue[store_idx].canWB = true;
694
695 ++storesToWB;
696 }
697
698 return checkViolations(load_idx, store_inst);
699
700}
701
702template <class Impl>
703void
704LSQUnit<Impl>::commitLoad()
705{
706 assert(loadQueue[loadHead]);
707
708 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n",
709 loadQueue[loadHead]->pcState());
710
711 loadQueue[loadHead] = NULL;
712
713 incrLdIdx(loadHead);
714
715 --loads;
716}
717
718template <class Impl>
719void
720LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst)
721{
722 assert(loads == 0 || loadQueue[loadHead]);
723
724 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) {
725 commitLoad();
726 }
727}
728
729template <class Impl>
730void
731LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst)
732{
733 assert(stores == 0 || storeQueue[storeHead].inst);
734
735 int store_idx = storeHead;
736
737 while (store_idx != storeTail) {
738 assert(storeQueue[store_idx].inst);
739 // Mark any stores that are now committed and have not yet
740 // been marked as able to write back.
741 if (!storeQueue[store_idx].canWB) {
742 if (storeQueue[store_idx].inst->seqNum > youngest_inst) {
743 break;
744 }
745 DPRINTF(LSQUnit, "Marking store as able to write back, PC "
746 "%s [sn:%lli]\n",
747 storeQueue[store_idx].inst->pcState(),
748 storeQueue[store_idx].inst->seqNum);
749
750 storeQueue[store_idx].canWB = true;
751
752 ++storesToWB;
753 }
754
755 incrStIdx(store_idx);
756 }
757}
758
759template <class Impl>
760void
761LSQUnit<Impl>::writebackPendingStore()
762{
763 if (hasPendingPkt) {
764 assert(pendingPkt != NULL);
765
766 // If the cache is blocked, this will store the packet for retry.
767 if (sendStore(pendingPkt)) {
768 storePostSend(pendingPkt);
769 }
770 pendingPkt = NULL;
771 hasPendingPkt = false;
772 }
773}
774
775template <class Impl>
776void
777LSQUnit<Impl>::writebackStores()
778{
779 // First writeback the second packet from any split store that didn't
780 // complete last cycle because there weren't enough cache ports available.
781 if (TheISA::HasUnalignedMemAcc) {
782 writebackPendingStore();
783 }
784
785 while (storesToWB > 0 &&
786 storeWBIdx != storeTail &&
787 storeQueue[storeWBIdx].inst &&
788 storeQueue[storeWBIdx].canWB &&
789 ((!needsTSO) || (!storeInFlight)) &&
790 usedPorts < cachePorts) {
791
792 if (isStoreBlocked || lsq->cacheBlocked()) {
793 DPRINTF(LSQUnit, "Unable to write back any more stores, cache"
794 " is blocked!\n");
795 break;
796 }
797
798 // Store didn't write any data so no need to write it back to
799 // memory.
800 if (storeQueue[storeWBIdx].size == 0) {
801 completeStore(storeWBIdx);
802
803 incrStIdx(storeWBIdx);
804
805 continue;
806 }
807
808 ++usedPorts;
809
810 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) {
811 incrStIdx(storeWBIdx);
812
813 continue;
814 }
815
816 assert(storeQueue[storeWBIdx].req);
817 assert(!storeQueue[storeWBIdx].committed);
818
819 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) {
820 assert(storeQueue[storeWBIdx].sreqLow);
821 assert(storeQueue[storeWBIdx].sreqHigh);
822 }
823
824 DynInstPtr inst = storeQueue[storeWBIdx].inst;
825
826 Request *req = storeQueue[storeWBIdx].req;
827 RequestPtr sreqLow = storeQueue[storeWBIdx].sreqLow;
828 RequestPtr sreqHigh = storeQueue[storeWBIdx].sreqHigh;
829
830 storeQueue[storeWBIdx].committed = true;
831
832 assert(!inst->memData);
833 inst->memData = new uint8_t[req->getSize()];
834
835 if (storeQueue[storeWBIdx].isAllZeros)
836 memset(inst->memData, 0, req->getSize());
837 else
838 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize());
839
840 MemCmd command =
841 req->isSwap() ? MemCmd::SwapReq :
842 (req->isLLSC() ? MemCmd::StoreCondReq : MemCmd::WriteReq);
843 PacketPtr data_pkt;
844 PacketPtr snd_data_pkt = NULL;
845
846 LSQSenderState *state = new LSQSenderState;
847 state->isLoad = false;
848 state->idx = storeWBIdx;
849 state->inst = inst;
850
851 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) {
852
853 // Build a single data packet if the store isn't split.
854 data_pkt = new Packet(req, command);
855 data_pkt->dataStatic(inst->memData);
856 data_pkt->senderState = state;
857 } else {
858 // Create two packets if the store is split in two.
859 data_pkt = new Packet(sreqLow, command);
860 snd_data_pkt = new Packet(sreqHigh, command);
861
862 data_pkt->dataStatic(inst->memData);
863 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize());
864
865 data_pkt->senderState = state;
866 snd_data_pkt->senderState = state;
867
868 state->isSplit = true;
869 state->outstanding = 2;
870
871 // Can delete the main request now.
872 delete req;
873 req = sreqLow;
874 }
875
876 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s "
877 "to Addr:%#x, data:%#x [sn:%lli]\n",
878 storeWBIdx, inst->pcState(),
879 req->getPaddr(), (int)*(inst->memData),
880 inst->seqNum);
881
882 // @todo: Remove this SC hack once the memory system handles it.
883 if (inst->isStoreConditional()) {
884 assert(!storeQueue[storeWBIdx].isSplit);
885 // Disable recording the result temporarily. Writing to
886 // misc regs normally updates the result, but this is not
887 // the desired behavior when handling store conditionals.
888 inst->recordResult(false);
889 bool success = TheISA::handleLockedWrite(inst.get(), req, cacheBlockMask);
890 inst->recordResult(true);
891
892 if (!success) {
893 // Instantly complete this store.
894 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. "
895 "Instantly completing it.\n",
896 inst->seqNum);
897 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this);
898 cpu->schedule(wb, curTick() + 1);
899 if (cpu->checker) {
900 // Make sure to set the LLSC data for verification
901 // if checker is loaded
902 inst->reqToVerify->setExtraData(0);
903 inst->completeAcc(data_pkt);
904 }
905 completeStore(storeWBIdx);
906 incrStIdx(storeWBIdx);
907 continue;
908 }
909 } else {
910 // Non-store conditionals do not need a writeback.
911 state->noWB = true;
912 }
913
914 bool split =
915 TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit;
916
917 ThreadContext *thread = cpu->tcBase(lsqID);
918
919 if (req->isMmappedIpr()) {
920 assert(!inst->isStoreConditional());
921 TheISA::handleIprWrite(thread, data_pkt);
922 delete data_pkt;
923 if (split) {
924 assert(snd_data_pkt->req->isMmappedIpr());
925 TheISA::handleIprWrite(thread, snd_data_pkt);
926 delete snd_data_pkt;
927 delete sreqLow;
928 delete sreqHigh;
929 }
930 delete state;
931 delete req;
932 completeStore(storeWBIdx);
933 incrStIdx(storeWBIdx);
934 } else if (!sendStore(data_pkt)) {
935 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will"
936 "retry later\n",
937 inst->seqNum);
938
939 // Need to store the second packet, if split.
940 if (split) {
941 state->pktToSend = true;
942 state->pendingPacket = snd_data_pkt;
943 }
944 } else {
945
946 // If split, try to send the second packet too
947 if (split) {
948 assert(snd_data_pkt);
949
950 // Ensure there are enough ports to use.
951 if (usedPorts < cachePorts) {
952 ++usedPorts;
953 if (sendStore(snd_data_pkt)) {
954 storePostSend(snd_data_pkt);
955 } else {
956 DPRINTF(IEW, "D-Cache became blocked when writing"
957 " [sn:%lli] second packet, will retry later\n",
958 inst->seqNum);
959 }
960 } else {
961
962 // Store the packet for when there's free ports.
963 assert(pendingPkt == NULL);
964 pendingPkt = snd_data_pkt;
965 hasPendingPkt = true;
966 }
967 } else {
968
969 // Not a split store.
970 storePostSend(data_pkt);
971 }
972 }
973 }
974
975 // Not sure this should set it to 0.
976 usedPorts = 0;
977
978 assert(stores >= 0 && storesToWB >= 0);
979}
980
981/*template <class Impl>
982void
983LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum)
984{
985 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(),
986 mshrSeqNums.end(),
987 seqNum);
988
989 if (mshr_it != mshrSeqNums.end()) {
990 mshrSeqNums.erase(mshr_it);
991 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size());
992 }
993}*/
994
995template <class Impl>
996void
997LSQUnit<Impl>::squash(const InstSeqNum &squashed_num)
998{
999 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!"
1000 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores);
1001
1002 int load_idx = loadTail;
1003 decrLdIdx(load_idx);
1004
1005 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) {
1006 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, "
1007 "[sn:%lli]\n",
1008 loadQueue[load_idx]->pcState(),
1009 loadQueue[load_idx]->seqNum);
1010
1011 if (isStalled() && load_idx == stallingLoadIdx) {
1012 stalled = false;
1013 stallingStoreIsn = 0;
1014 stallingLoadIdx = 0;
1015 }
1016
1017 // Clear the smart pointer to make sure it is decremented.
1018 loadQueue[load_idx]->setSquashed();
1019 loadQueue[load_idx] = NULL;
1020 --loads;
1021
1022 // Inefficient!
1023 loadTail = load_idx;
1024
1025 decrLdIdx(load_idx);
1026 ++lsqSquashedLoads;
1027 }
1028
1029 if (isLoadBlocked) {
1030 if (squashed_num < blockedLoadSeqNum) {
1031 isLoadBlocked = false;
1032 loadBlockedHandled = false;
1033 blockedLoadSeqNum = 0;
1034 }
1035 }
1036
1037 if (memDepViolator && squashed_num < memDepViolator->seqNum) {
1038 memDepViolator = NULL;
1039 }
1040
1041 int store_idx = storeTail;
1042 decrStIdx(store_idx);
1043
1044 while (stores != 0 &&
1045 storeQueue[store_idx].inst->seqNum > squashed_num) {
1046 // Instructions marked as can WB are already committed.
1047 if (storeQueue[store_idx].canWB) {
1048 break;
1049 }
1050
1051 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, "
1052 "idx:%i [sn:%lli]\n",
1053 storeQueue[store_idx].inst->pcState(),
1054 store_idx, storeQueue[store_idx].inst->seqNum);
1055
1056 // I don't think this can happen. It should have been cleared
1057 // by the stalling load.
1058 if (isStalled() &&
1059 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) {
1060 panic("Is stalled should have been cleared by stalling load!\n");
1061 stalled = false;
1062 stallingStoreIsn = 0;
1063 }
1064
1065 // Clear the smart pointer to make sure it is decremented.
1066 storeQueue[store_idx].inst->setSquashed();
1067 storeQueue[store_idx].inst = NULL;
1068 storeQueue[store_idx].canWB = 0;
1069
1070 // Must delete request now that it wasn't handed off to
1071 // memory. This is quite ugly. @todo: Figure out the proper
1072 // place to really handle request deletes.
1073 delete storeQueue[store_idx].req;
1074 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) {
1075 delete storeQueue[store_idx].sreqLow;
1076 delete storeQueue[store_idx].sreqHigh;
1077
1078 storeQueue[store_idx].sreqLow = NULL;
1079 storeQueue[store_idx].sreqHigh = NULL;
1080 }
1081
1082 storeQueue[store_idx].req = NULL;
1083 --stores;
1084
1085 // Inefficient!
1086 storeTail = store_idx;
1087
1088 decrStIdx(store_idx);
1089 ++lsqSquashedStores;
1090 }
1091}
1092
1093template <class Impl>
1094void
1095LSQUnit<Impl>::storePostSend(PacketPtr pkt)
1096{
1097 if (isStalled() &&
1098 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) {
1099 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] "
1100 "load idx:%i\n",
1101 stallingStoreIsn, stallingLoadIdx);
1102 stalled = false;
1103 stallingStoreIsn = 0;
1104 iewStage->replayMemInst(loadQueue[stallingLoadIdx]);
1105 }
1106
1107 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) {
1108 // The store is basically completed at this time. This
1109 // only works so long as the checker doesn't try to
1110 // verify the value in memory for stores.
1111 storeQueue[storeWBIdx].inst->setCompleted();
1112
1113 if (cpu->checker) {
1114 cpu->checker->verify(storeQueue[storeWBIdx].inst);
1115 }
1116 }
1117
1118 if (needsTSO) {
1119 storeInFlight = true;
1120 }
1121
1122 incrStIdx(storeWBIdx);
1123}
1124
1125template <class Impl>
1126void
1127LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt)
1128{
1129 iewStage->wakeCPU();
1130
1131 // Squashed instructions do not need to complete their access.
1132 if (inst->isSquashed()) {
1133 iewStage->decrWb(inst->seqNum);
1134 assert(!inst->isStore());
1135 ++lsqIgnoredResponses;
1136 return;
1137 }
1138
1139 if (!inst->isExecuted()) {
1140 inst->setExecuted();
1141
1142 // Complete access to copy data to proper place.
1143 inst->completeAcc(pkt);
1144 }
1145
1146 // Need to insert instruction into queue to commit
1147 iewStage->instToCommit(inst);
1148
1149 iewStage->activityThisCycle();
1150
1151 // see if this load changed the PC
1152 iewStage->checkMisprediction(inst);
1153}
1154
1155template <class Impl>
1156void
1157LSQUnit<Impl>::completeStore(int store_idx)
1158{
1159 assert(storeQueue[store_idx].inst);
1160 storeQueue[store_idx].completed = true;
1161 --storesToWB;
1162 // A bit conservative because a store completion may not free up entries,
1163 // but hopefully avoids two store completions in one cycle from making
1164 // the CPU tick twice.
1165 cpu->wakeCPU();
1166 cpu->activityThisCycle();
1167
1168 if (store_idx == storeHead) {
1169 do {
1170 incrStIdx(storeHead);
1171
1172 --stores;
1173 } while (storeQueue[storeHead].completed &&
1174 storeHead != storeTail);
1175
1176 iewStage->updateLSQNextCycle = true;
1177 }
1178
1179 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head "
1180 "idx:%i\n",
1181 storeQueue[store_idx].inst->seqNum, store_idx, storeHead);
1182
1183#if TRACING_ON
1184 if (DTRACE(O3PipeView)) {
1185 storeQueue[store_idx].inst->storeTick =
1186 curTick() - storeQueue[store_idx].inst->fetchTick;
1187 }
1188#endif
1189
1190 if (isStalled() &&
1191 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) {
1192 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] "
1193 "load idx:%i\n",
1194 stallingStoreIsn, stallingLoadIdx);
1195 stalled = false;
1196 stallingStoreIsn = 0;
1197 iewStage->replayMemInst(loadQueue[stallingLoadIdx]);
1198 }
1199
1200 storeQueue[store_idx].inst->setCompleted();
1201
1202 if (needsTSO) {
1203 storeInFlight = false;
1204 }
1205
1206 // Tell the checker we've completed this instruction. Some stores
1207 // may get reported twice to the checker, but the checker can
1208 // handle that case.
1209 if (cpu->checker) {
1210 cpu->checker->verify(storeQueue[store_idx].inst);
1211 }
1212}
1213
1214template <class Impl>
1215bool
1216LSQUnit<Impl>::sendStore(PacketPtr data_pkt)
1217{
1218 if (!dcachePort->sendTimingReq(data_pkt)) {
1219 // Need to handle becoming blocked on a store.
1220 isStoreBlocked = true;
1221 ++lsqCacheBlocked;
1222 assert(retryPkt == NULL);
1223 retryPkt = data_pkt;
1224 lsq->setRetryTid(lsqID);
1225 return false;
1226 }
1227 return true;
1228}
1229
1230template <class Impl>
1231void
1232LSQUnit<Impl>::recvRetry()
1233{
1234 if (isStoreBlocked) {
1235 DPRINTF(LSQUnit, "Receiving retry: store blocked\n");
1236 assert(retryPkt != NULL);
1237
1238 LSQSenderState *state =
1239 dynamic_cast<LSQSenderState *>(retryPkt->senderState);
1240
1241 if (dcachePort->sendTimingReq(retryPkt)) {
1242 // Don't finish the store unless this is the last packet.
1243 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend ||
1244 state->pendingPacket == retryPkt) {
1245 state->pktToSend = false;
1246 storePostSend(retryPkt);
1247 }
1248 retryPkt = NULL;
1249 isStoreBlocked = false;
1250 lsq->setRetryTid(InvalidThreadID);
1251
1252 // Send any outstanding packet.
1253 if (TheISA::HasUnalignedMemAcc && state->pktToSend) {
1254 assert(state->pendingPacket);
1255 if (sendStore(state->pendingPacket)) {
1256 storePostSend(state->pendingPacket);
1257 }
1258 }
1259 } else {
1260 // Still blocked!
1261 ++lsqCacheBlocked;
1262 lsq->setRetryTid(lsqID);
1263 }
1264 } else if (isLoadBlocked) {
1265 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, "
1266 "no need to resend packet.\n");
1267 } else {
1268 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n");
1269 }
1270}
1271
1272template <class Impl>
1273inline void
1274LSQUnit<Impl>::incrStIdx(int &store_idx) const
1275{
1276 if (++store_idx >= SQEntries)
1277 store_idx = 0;
1278}
1279
1280template <class Impl>
1281inline void
1282LSQUnit<Impl>::decrStIdx(int &store_idx) const
1283{
1284 if (--store_idx < 0)
1285 store_idx += SQEntries;
1286}
1287
1288template <class Impl>
1289inline void
1290LSQUnit<Impl>::incrLdIdx(int &load_idx) const
1291{
1292 if (++load_idx >= LQEntries)
1293 load_idx = 0;
1294}
1295
1296template <class Impl>
1297inline void
1298LSQUnit<Impl>::decrLdIdx(int &load_idx) const
1299{
1300 if (--load_idx < 0)
1301 load_idx += LQEntries;
1302}
1303
1304template <class Impl>
1305void
1306LSQUnit<Impl>::dumpInsts() const
1307{
1308 cprintf("Load store queue: Dumping instructions.\n");
1309 cprintf("Load queue size: %i\n", loads);
1310 cprintf("Load queue: ");
1311
1312 int load_idx = loadHead;
1313
1314 while (load_idx != loadTail && loadQueue[load_idx]) {
1315 const DynInstPtr &inst(loadQueue[load_idx]);
1316 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum);
1317
1318 incrLdIdx(load_idx);
1319 }
1320 cprintf("\n");
1321
1322 cprintf("Store queue size: %i\n", stores);
1323 cprintf("Store queue: ");
1324
1325 int store_idx = storeHead;
1326
1327 while (store_idx != storeTail && storeQueue[store_idx].inst) {
1328 const DynInstPtr &inst(storeQueue[store_idx].inst);
1329 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum);
1330
1331 incrStIdx(store_idx);
1332 }
1333
1334 cprintf("\n");
1335}
1336
1337#endif//__CPU_O3_LSQ_UNIT_IMPL_HH__