Deleted Added
sdiff udiff text old ( 10240:15f822e9410a ) new ( 10327:5b6279635c49 )
full compact
1/*
2 * Copyright (c) 2010-2013 ARM Limited
3 * Copyright (c) 2013 Advanced Micro Devices, Inc.
4 * All rights reserved.
5 *
6 * The license below extends only to copyright in the software and shall
7 * not be construed as granting a license to any other intellectual
8 * property including but not limited to intellectual property relating
9 * to a hardware implementation of the functionality of the software
10 * licensed hereunder. You may use the software subject to the license
11 * terms below provided that you ensure that this notice is replicated
12 * unmodified and in its entirety in all distributions of the software,
13 * modified or unmodified, in source code or in binary form.
14 *
15 * Copyright (c) 2004-2006 The Regents of The University of Michigan
16 * All rights reserved.
17 *
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions are
20 * met: redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer;
22 * redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution;
25 * neither the name of the copyright holders nor the names of its
26 * contributors may be used to endorse or promote products derived from
27 * this software without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 *
41 * Authors: Kevin Lim
42 */
43
44#ifndef __CPU_O3_IEW_IMPL_IMPL_HH__
45#define __CPU_O3_IEW_IMPL_IMPL_HH__
46
47// @todo: Fix the instantaneous communication among all the stages within
48// iew. There's a clear delay between issue and execute, yet backwards
49// communication happens simultaneously.
50
51#include <queue>
52
53#include "arch/utility.hh"
54#include "config/the_isa.hh"
55#include "cpu/checker/cpu.hh"
56#include "cpu/o3/fu_pool.hh"
57#include "cpu/o3/iew.hh"
58#include "cpu/timebuf.hh"
59#include "debug/Activity.hh"
60#include "debug/Drain.hh"
61#include "debug/IEW.hh"
62#include "debug/O3PipeView.hh"
63#include "params/DerivO3CPU.hh"
64
65using namespace std;
66
67template<class Impl>
68DefaultIEW<Impl>::DefaultIEW(O3CPU *_cpu, DerivO3CPUParams *params)
69 : issueToExecQueue(params->backComSize, params->forwardComSize),
70 cpu(_cpu),
71 instQueue(_cpu, this, params),
72 ldstQueue(_cpu, this, params),
73 fuPool(params->fuPool),
74 commitToIEWDelay(params->commitToIEWDelay),
75 renameToIEWDelay(params->renameToIEWDelay),
76 issueToExecuteDelay(params->issueToExecuteDelay),
77 dispatchWidth(params->dispatchWidth),
78 issueWidth(params->issueWidth),
79 wbWidth(params->wbWidth),
80 numThreads(params->numThreads)
81{
82 if (dispatchWidth > Impl::MaxWidth)
83 fatal("dispatchWidth (%d) is larger than compiled limit (%d),\n"
84 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
85 dispatchWidth, static_cast<int>(Impl::MaxWidth));
86 if (issueWidth > Impl::MaxWidth)
87 fatal("issueWidth (%d) is larger than compiled limit (%d),\n"
88 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
89 issueWidth, static_cast<int>(Impl::MaxWidth));
90 if (wbWidth > Impl::MaxWidth)
91 fatal("wbWidth (%d) is larger than compiled limit (%d),\n"
92 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
93 wbWidth, static_cast<int>(Impl::MaxWidth));
94
95 _status = Active;
96 exeStatus = Running;
97 wbStatus = Idle;
98
99 // Setup wire to read instructions coming from issue.
100 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
101
102 // Instruction queue needs the queue between issue and execute.
103 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
104
105 for (ThreadID tid = 0; tid < numThreads; tid++) {
106 dispatchStatus[tid] = Running;
107 stalls[tid].commit = false;
108 fetchRedirect[tid] = false;
109 }
110
111 updateLSQNextCycle = false;
112
113 skidBufferMax = (3 * (renameToIEWDelay * params->renameWidth)) + issueWidth;
114}
115
116template <class Impl>
117std::string
118DefaultIEW<Impl>::name() const
119{
120 return cpu->name() + ".iew";
121}
122
123template <class Impl>
124void
125DefaultIEW<Impl>::regProbePoints()
126{
127 ppDispatch = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Dispatch");
128 ppMispredict = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Mispredict");
129}
130
131template <class Impl>
132void
133DefaultIEW<Impl>::regStats()
134{
135 using namespace Stats;
136
137 instQueue.regStats();
138 ldstQueue.regStats();
139
140 iewIdleCycles
141 .name(name() + ".iewIdleCycles")
142 .desc("Number of cycles IEW is idle");
143
144 iewSquashCycles
145 .name(name() + ".iewSquashCycles")
146 .desc("Number of cycles IEW is squashing");
147
148 iewBlockCycles
149 .name(name() + ".iewBlockCycles")
150 .desc("Number of cycles IEW is blocking");
151
152 iewUnblockCycles
153 .name(name() + ".iewUnblockCycles")
154 .desc("Number of cycles IEW is unblocking");
155
156 iewDispatchedInsts
157 .name(name() + ".iewDispatchedInsts")
158 .desc("Number of instructions dispatched to IQ");
159
160 iewDispSquashedInsts
161 .name(name() + ".iewDispSquashedInsts")
162 .desc("Number of squashed instructions skipped by dispatch");
163
164 iewDispLoadInsts
165 .name(name() + ".iewDispLoadInsts")
166 .desc("Number of dispatched load instructions");
167
168 iewDispStoreInsts
169 .name(name() + ".iewDispStoreInsts")
170 .desc("Number of dispatched store instructions");
171
172 iewDispNonSpecInsts
173 .name(name() + ".iewDispNonSpecInsts")
174 .desc("Number of dispatched non-speculative instructions");
175
176 iewIQFullEvents
177 .name(name() + ".iewIQFullEvents")
178 .desc("Number of times the IQ has become full, causing a stall");
179
180 iewLSQFullEvents
181 .name(name() + ".iewLSQFullEvents")
182 .desc("Number of times the LSQ has become full, causing a stall");
183
184 memOrderViolationEvents
185 .name(name() + ".memOrderViolationEvents")
186 .desc("Number of memory order violations");
187
188 predictedTakenIncorrect
189 .name(name() + ".predictedTakenIncorrect")
190 .desc("Number of branches that were predicted taken incorrectly");
191
192 predictedNotTakenIncorrect
193 .name(name() + ".predictedNotTakenIncorrect")
194 .desc("Number of branches that were predicted not taken incorrectly");
195
196 branchMispredicts
197 .name(name() + ".branchMispredicts")
198 .desc("Number of branch mispredicts detected at execute");
199
200 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
201
202 iewExecutedInsts
203 .name(name() + ".iewExecutedInsts")
204 .desc("Number of executed instructions");
205
206 iewExecLoadInsts
207 .init(cpu->numThreads)
208 .name(name() + ".iewExecLoadInsts")
209 .desc("Number of load instructions executed")
210 .flags(total);
211
212 iewExecSquashedInsts
213 .name(name() + ".iewExecSquashedInsts")
214 .desc("Number of squashed instructions skipped in execute");
215
216 iewExecutedSwp
217 .init(cpu->numThreads)
218 .name(name() + ".exec_swp")
219 .desc("number of swp insts executed")
220 .flags(total);
221
222 iewExecutedNop
223 .init(cpu->numThreads)
224 .name(name() + ".exec_nop")
225 .desc("number of nop insts executed")
226 .flags(total);
227
228 iewExecutedRefs
229 .init(cpu->numThreads)
230 .name(name() + ".exec_refs")
231 .desc("number of memory reference insts executed")
232 .flags(total);
233
234 iewExecutedBranches
235 .init(cpu->numThreads)
236 .name(name() + ".exec_branches")
237 .desc("Number of branches executed")
238 .flags(total);
239
240 iewExecStoreInsts
241 .name(name() + ".exec_stores")
242 .desc("Number of stores executed")
243 .flags(total);
244 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
245
246 iewExecRate
247 .name(name() + ".exec_rate")
248 .desc("Inst execution rate")
249 .flags(total);
250
251 iewExecRate = iewExecutedInsts / cpu->numCycles;
252
253 iewInstsToCommit
254 .init(cpu->numThreads)
255 .name(name() + ".wb_sent")
256 .desc("cumulative count of insts sent to commit")
257 .flags(total);
258
259 writebackCount
260 .init(cpu->numThreads)
261 .name(name() + ".wb_count")
262 .desc("cumulative count of insts written-back")
263 .flags(total);
264
265 producerInst
266 .init(cpu->numThreads)
267 .name(name() + ".wb_producers")
268 .desc("num instructions producing a value")
269 .flags(total);
270
271 consumerInst
272 .init(cpu->numThreads)
273 .name(name() + ".wb_consumers")
274 .desc("num instructions consuming a value")
275 .flags(total);
276
277 wbPenalized
278 .init(cpu->numThreads)
279 .name(name() + ".wb_penalized")
280 .desc("number of instrctions required to write to 'other' IQ")
281 .flags(total);
282
283 wbPenalizedRate
284 .name(name() + ".wb_penalized_rate")
285 .desc ("fraction of instructions written-back that wrote to 'other' IQ")
286 .flags(total);
287
288 wbPenalizedRate = wbPenalized / writebackCount;
289
290 wbFanout
291 .name(name() + ".wb_fanout")
292 .desc("average fanout of values written-back")
293 .flags(total);
294
295 wbFanout = producerInst / consumerInst;
296
297 wbRate
298 .name(name() + ".wb_rate")
299 .desc("insts written-back per cycle")
300 .flags(total);
301 wbRate = writebackCount / cpu->numCycles;
302}
303
304template<class Impl>
305void
306DefaultIEW<Impl>::startupStage()
307{
308 for (ThreadID tid = 0; tid < numThreads; tid++) {
309 toRename->iewInfo[tid].usedIQ = true;
310 toRename->iewInfo[tid].freeIQEntries =
311 instQueue.numFreeEntries(tid);
312
313 toRename->iewInfo[tid].usedLSQ = true;
314 toRename->iewInfo[tid].freeLQEntries = ldstQueue.numFreeLoadEntries(tid);
315 toRename->iewInfo[tid].freeSQEntries = ldstQueue.numFreeStoreEntries(tid);
316 }
317
318 // Initialize the checker's dcache port here
319 if (cpu->checker) {
320 cpu->checker->setDcachePort(&cpu->getDataPort());
321 }
322
323 cpu->activateStage(O3CPU::IEWIdx);
324}
325
326template<class Impl>
327void
328DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
329{
330 timeBuffer = tb_ptr;
331
332 // Setup wire to read information from time buffer, from commit.
333 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
334
335 // Setup wire to write information back to previous stages.
336 toRename = timeBuffer->getWire(0);
337
338 toFetch = timeBuffer->getWire(0);
339
340 // Instruction queue also needs main time buffer.
341 instQueue.setTimeBuffer(tb_ptr);
342}
343
344template<class Impl>
345void
346DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
347{
348 renameQueue = rq_ptr;
349
350 // Setup wire to read information from rename queue.
351 fromRename = renameQueue->getWire(-renameToIEWDelay);
352}
353
354template<class Impl>
355void
356DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
357{
358 iewQueue = iq_ptr;
359
360 // Setup wire to write instructions to commit.
361 toCommit = iewQueue->getWire(0);
362}
363
364template<class Impl>
365void
366DefaultIEW<Impl>::setActiveThreads(list<ThreadID> *at_ptr)
367{
368 activeThreads = at_ptr;
369
370 ldstQueue.setActiveThreads(at_ptr);
371 instQueue.setActiveThreads(at_ptr);
372}
373
374template<class Impl>
375void
376DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
377{
378 scoreboard = sb_ptr;
379}
380
381template <class Impl>
382bool
383DefaultIEW<Impl>::isDrained() const
384{
385 bool drained(ldstQueue.isDrained());
386
387 for (ThreadID tid = 0; tid < numThreads; tid++) {
388 if (!insts[tid].empty()) {
389 DPRINTF(Drain, "%i: Insts not empty.\n", tid);
390 drained = false;
391 }
392 if (!skidBuffer[tid].empty()) {
393 DPRINTF(Drain, "%i: Skid buffer not empty.\n", tid);
394 drained = false;
395 }
396 }
397
398 // Also check the FU pool as instructions are "stored" in FU
399 // completion events until they are done and not accounted for
400 // above
401 if (drained && !fuPool->isDrained()) {
402 DPRINTF(Drain, "FU pool still busy.\n");
403 drained = false;
404 }
405
406 return drained;
407}
408
409template <class Impl>
410void
411DefaultIEW<Impl>::drainSanityCheck() const
412{
413 assert(isDrained());
414
415 instQueue.drainSanityCheck();
416 ldstQueue.drainSanityCheck();
417}
418
419template <class Impl>
420void
421DefaultIEW<Impl>::takeOverFrom()
422{
423 // Reset all state.
424 _status = Active;
425 exeStatus = Running;
426 wbStatus = Idle;
427
428 instQueue.takeOverFrom();
429 ldstQueue.takeOverFrom();
430 fuPool->takeOverFrom();
431
432 startupStage();
433 cpu->activityThisCycle();
434
435 for (ThreadID tid = 0; tid < numThreads; tid++) {
436 dispatchStatus[tid] = Running;
437 stalls[tid].commit = false;
438 fetchRedirect[tid] = false;
439 }
440
441 updateLSQNextCycle = false;
442
443 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
444 issueToExecQueue.advance();
445 }
446}
447
448template<class Impl>
449void
450DefaultIEW<Impl>::squash(ThreadID tid)
451{
452 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n", tid);
453
454 // Tell the IQ to start squashing.
455 instQueue.squash(tid);
456
457 // Tell the LDSTQ to start squashing.
458 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
459 updatedQueues = true;
460
461 // Clear the skid buffer in case it has any data in it.
462 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
463 tid, fromCommit->commitInfo[tid].doneSeqNum);
464
465 while (!skidBuffer[tid].empty()) {
466 if (skidBuffer[tid].front()->isLoad()) {
467 toRename->iewInfo[tid].dispatchedToLQ++;
468 }
469 if (skidBuffer[tid].front()->isStore()) {
470 toRename->iewInfo[tid].dispatchedToSQ++;
471 }
472
473 toRename->iewInfo[tid].dispatched++;
474
475 skidBuffer[tid].pop();
476 }
477
478 emptyRenameInsts(tid);
479}
480
481template<class Impl>
482void
483DefaultIEW<Impl>::squashDueToBranch(DynInstPtr &inst, ThreadID tid)
484{
485 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %s "
486 "[sn:%i].\n", tid, inst->pcState(), inst->seqNum);
487
488 if (!toCommit->squash[tid] ||
489 inst->seqNum < toCommit->squashedSeqNum[tid]) {
490 toCommit->squash[tid] = true;
491 toCommit->squashedSeqNum[tid] = inst->seqNum;
492 toCommit->branchTaken[tid] = inst->pcState().branching();
493
494 TheISA::PCState pc = inst->pcState();
495 TheISA::advancePC(pc, inst->staticInst);
496
497 toCommit->pc[tid] = pc;
498 toCommit->mispredictInst[tid] = inst;
499 toCommit->includeSquashInst[tid] = false;
500
501 wroteToTimeBuffer = true;
502 }
503
504}
505
506template<class Impl>
507void
508DefaultIEW<Impl>::squashDueToMemOrder(DynInstPtr &inst, ThreadID tid)
509{
510 DPRINTF(IEW, "[tid:%i]: Memory violation, squashing violator and younger "
511 "insts, PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
512 // Need to include inst->seqNum in the following comparison to cover the
513 // corner case when a branch misprediction and a memory violation for the
514 // same instruction (e.g. load PC) are detected in the same cycle. In this
515 // case the memory violator should take precedence over the branch
516 // misprediction because it requires the violator itself to be included in
517 // the squash.
518 if (!toCommit->squash[tid] ||
519 inst->seqNum <= toCommit->squashedSeqNum[tid]) {
520 toCommit->squash[tid] = true;
521
522 toCommit->squashedSeqNum[tid] = inst->seqNum;
523 toCommit->pc[tid] = inst->pcState();
524 toCommit->mispredictInst[tid] = NULL;
525
526 // Must include the memory violator in the squash.
527 toCommit->includeSquashInst[tid] = true;
528
529 wroteToTimeBuffer = true;
530 }
531}
532
533template<class Impl>
534void
535DefaultIEW<Impl>::squashDueToMemBlocked(DynInstPtr &inst, ThreadID tid)
536{
537 DPRINTF(IEW, "[tid:%i]: Memory blocked, squashing load and younger insts, "
538 "PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
539 if (!toCommit->squash[tid] ||
540 inst->seqNum < toCommit->squashedSeqNum[tid]) {
541 toCommit->squash[tid] = true;
542
543 toCommit->squashedSeqNum[tid] = inst->seqNum;
544 toCommit->pc[tid] = inst->pcState();
545 toCommit->mispredictInst[tid] = NULL;
546
547 // Must include the broadcasted SN in the squash.
548 toCommit->includeSquashInst[tid] = true;
549
550 ldstQueue.setLoadBlockedHandled(tid);
551
552 wroteToTimeBuffer = true;
553 }
554}
555
556template<class Impl>
557void
558DefaultIEW<Impl>::block(ThreadID tid)
559{
560 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
561
562 if (dispatchStatus[tid] != Blocked &&
563 dispatchStatus[tid] != Unblocking) {
564 toRename->iewBlock[tid] = true;
565 wroteToTimeBuffer = true;
566 }
567
568 // Add the current inputs to the skid buffer so they can be
569 // reprocessed when this stage unblocks.
570 skidInsert(tid);
571
572 dispatchStatus[tid] = Blocked;
573}
574
575template<class Impl>
576void
577DefaultIEW<Impl>::unblock(ThreadID tid)
578{
579 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
580 "buffer %u.\n",tid, tid);
581
582 // If the skid bufffer is empty, signal back to previous stages to unblock.
583 // Also switch status to running.
584 if (skidBuffer[tid].empty()) {
585 toRename->iewUnblock[tid] = true;
586 wroteToTimeBuffer = true;
587 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
588 dispatchStatus[tid] = Running;
589 }
590}
591
592template<class Impl>
593void
594DefaultIEW<Impl>::wakeDependents(DynInstPtr &inst)
595{
596 instQueue.wakeDependents(inst);
597}
598
599template<class Impl>
600void
601DefaultIEW<Impl>::rescheduleMemInst(DynInstPtr &inst)
602{
603 instQueue.rescheduleMemInst(inst);
604}
605
606template<class Impl>
607void
608DefaultIEW<Impl>::replayMemInst(DynInstPtr &inst)
609{
610 instQueue.replayMemInst(inst);
611}
612
613template<class Impl>
614void
615DefaultIEW<Impl>::instToCommit(DynInstPtr &inst)
616{
617 // This function should not be called after writebackInsts in a
618 // single cycle. That will cause problems with an instruction
619 // being added to the queue to commit without being processed by
620 // writebackInsts prior to being sent to commit.
621
622 // First check the time slot that this instruction will write
623 // to. If there are free write ports at the time, then go ahead
624 // and write the instruction to that time. If there are not,
625 // keep looking back to see where's the first time there's a
626 // free slot.
627 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
628 ++wbNumInst;
629 if (wbNumInst == wbWidth) {
630 ++wbCycle;
631 wbNumInst = 0;
632 }
633 }
634
635 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
636 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
637 // Add finished instruction to queue to commit.
638 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
639 (*iewQueue)[wbCycle].size++;
640}
641
642template <class Impl>
643unsigned
644DefaultIEW<Impl>::validInstsFromRename()
645{
646 unsigned inst_count = 0;
647
648 for (int i=0; i<fromRename->size; i++) {
649 if (!fromRename->insts[i]->isSquashed())
650 inst_count++;
651 }
652
653 return inst_count;
654}
655
656template<class Impl>
657void
658DefaultIEW<Impl>::skidInsert(ThreadID tid)
659{
660 DynInstPtr inst = NULL;
661
662 while (!insts[tid].empty()) {
663 inst = insts[tid].front();
664
665 insts[tid].pop();
666
667 DPRINTF(IEW,"[tid:%i]: Inserting [sn:%lli] PC:%s into "
668 "dispatch skidBuffer %i\n",tid, inst->seqNum,
669 inst->pcState(),tid);
670
671 skidBuffer[tid].push(inst);
672 }
673
674 assert(skidBuffer[tid].size() <= skidBufferMax &&
675 "Skidbuffer Exceeded Max Size");
676}
677
678template<class Impl>
679int
680DefaultIEW<Impl>::skidCount()
681{
682 int max=0;
683
684 list<ThreadID>::iterator threads = activeThreads->begin();
685 list<ThreadID>::iterator end = activeThreads->end();
686
687 while (threads != end) {
688 ThreadID tid = *threads++;
689 unsigned thread_count = skidBuffer[tid].size();
690 if (max < thread_count)
691 max = thread_count;
692 }
693
694 return max;
695}
696
697template<class Impl>
698bool
699DefaultIEW<Impl>::skidsEmpty()
700{
701 list<ThreadID>::iterator threads = activeThreads->begin();
702 list<ThreadID>::iterator end = activeThreads->end();
703
704 while (threads != end) {
705 ThreadID tid = *threads++;
706
707 if (!skidBuffer[tid].empty())
708 return false;
709 }
710
711 return true;
712}
713
714template <class Impl>
715void
716DefaultIEW<Impl>::updateStatus()
717{
718 bool any_unblocking = false;
719
720 list<ThreadID>::iterator threads = activeThreads->begin();
721 list<ThreadID>::iterator end = activeThreads->end();
722
723 while (threads != end) {
724 ThreadID tid = *threads++;
725
726 if (dispatchStatus[tid] == Unblocking) {
727 any_unblocking = true;
728 break;
729 }
730 }
731
732 // If there are no ready instructions waiting to be scheduled by the IQ,
733 // and there's no stores waiting to write back, and dispatch is not
734 // unblocking, then there is no internal activity for the IEW stage.
735 instQueue.intInstQueueReads++;
736 if (_status == Active && !instQueue.hasReadyInsts() &&
737 !ldstQueue.willWB() && !any_unblocking) {
738 DPRINTF(IEW, "IEW switching to idle\n");
739
740 deactivateStage();
741
742 _status = Inactive;
743 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
744 ldstQueue.willWB() ||
745 any_unblocking)) {
746 // Otherwise there is internal activity. Set to active.
747 DPRINTF(IEW, "IEW switching to active\n");
748
749 activateStage();
750
751 _status = Active;
752 }
753}
754
755template <class Impl>
756void
757DefaultIEW<Impl>::resetEntries()
758{
759 instQueue.resetEntries();
760 ldstQueue.resetEntries();
761}
762
763template <class Impl>
764void
765DefaultIEW<Impl>::readStallSignals(ThreadID tid)
766{
767 if (fromCommit->commitBlock[tid]) {
768 stalls[tid].commit = true;
769 }
770
771 if (fromCommit->commitUnblock[tid]) {
772 assert(stalls[tid].commit);
773 stalls[tid].commit = false;
774 }
775}
776
777template <class Impl>
778bool
779DefaultIEW<Impl>::checkStall(ThreadID tid)
780{
781 bool ret_val(false);
782
783 if (stalls[tid].commit) {
784 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
785 ret_val = true;
786 } else if (instQueue.isFull(tid)) {
787 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
788 ret_val = true;
789 }
790
791 return ret_val;
792}
793
794template <class Impl>
795void
796DefaultIEW<Impl>::checkSignalsAndUpdate(ThreadID tid)
797{
798 // Check if there's a squash signal, squash if there is
799 // Check stall signals, block if there is.
800 // If status was Blocked
801 // if so then go to unblocking
802 // If status was Squashing
803 // check if squashing is not high. Switch to running this cycle.
804
805 readStallSignals(tid);
806
807 if (fromCommit->commitInfo[tid].squash) {
808 squash(tid);
809
810 if (dispatchStatus[tid] == Blocked ||
811 dispatchStatus[tid] == Unblocking) {
812 toRename->iewUnblock[tid] = true;
813 wroteToTimeBuffer = true;
814 }
815
816 dispatchStatus[tid] = Squashing;
817 fetchRedirect[tid] = false;
818 return;
819 }
820
821 if (fromCommit->commitInfo[tid].robSquashing) {
822 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
823
824 dispatchStatus[tid] = Squashing;
825 emptyRenameInsts(tid);
826 wroteToTimeBuffer = true;
827 return;
828 }
829
830 if (checkStall(tid)) {
831 block(tid);
832 dispatchStatus[tid] = Blocked;
833 return;
834 }
835
836 if (dispatchStatus[tid] == Blocked) {
837 // Status from previous cycle was blocked, but there are no more stall
838 // conditions. Switch over to unblocking.
839 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
840 tid);
841
842 dispatchStatus[tid] = Unblocking;
843
844 unblock(tid);
845
846 return;
847 }
848
849 if (dispatchStatus[tid] == Squashing) {
850 // Switch status to running if rename isn't being told to block or
851 // squash this cycle.
852 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
853 tid);
854
855 dispatchStatus[tid] = Running;
856
857 return;
858 }
859}
860
861template <class Impl>
862void
863DefaultIEW<Impl>::sortInsts()
864{
865 int insts_from_rename = fromRename->size;
866#ifdef DEBUG
867 for (ThreadID tid = 0; tid < numThreads; tid++)
868 assert(insts[tid].empty());
869#endif
870 for (int i = 0; i < insts_from_rename; ++i) {
871 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
872 }
873}
874
875template <class Impl>
876void
877DefaultIEW<Impl>::emptyRenameInsts(ThreadID tid)
878{
879 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions\n", tid);
880
881 while (!insts[tid].empty()) {
882
883 if (insts[tid].front()->isLoad()) {
884 toRename->iewInfo[tid].dispatchedToLQ++;
885 }
886 if (insts[tid].front()->isStore()) {
887 toRename->iewInfo[tid].dispatchedToSQ++;
888 }
889
890 toRename->iewInfo[tid].dispatched++;
891
892 insts[tid].pop();
893 }
894}
895
896template <class Impl>
897void
898DefaultIEW<Impl>::wakeCPU()
899{
900 cpu->wakeCPU();
901}
902
903template <class Impl>
904void
905DefaultIEW<Impl>::activityThisCycle()
906{
907 DPRINTF(Activity, "Activity this cycle.\n");
908 cpu->activityThisCycle();
909}
910
911template <class Impl>
912inline void
913DefaultIEW<Impl>::activateStage()
914{
915 DPRINTF(Activity, "Activating stage.\n");
916 cpu->activateStage(O3CPU::IEWIdx);
917}
918
919template <class Impl>
920inline void
921DefaultIEW<Impl>::deactivateStage()
922{
923 DPRINTF(Activity, "Deactivating stage.\n");
924 cpu->deactivateStage(O3CPU::IEWIdx);
925}
926
927template<class Impl>
928void
929DefaultIEW<Impl>::dispatch(ThreadID tid)
930{
931 // If status is Running or idle,
932 // call dispatchInsts()
933 // If status is Unblocking,
934 // buffer any instructions coming from rename
935 // continue trying to empty skid buffer
936 // check if stall conditions have passed
937
938 if (dispatchStatus[tid] == Blocked) {
939 ++iewBlockCycles;
940
941 } else if (dispatchStatus[tid] == Squashing) {
942 ++iewSquashCycles;
943 }
944
945 // Dispatch should try to dispatch as many instructions as its bandwidth
946 // will allow, as long as it is not currently blocked.
947 if (dispatchStatus[tid] == Running ||
948 dispatchStatus[tid] == Idle) {
949 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
950 "dispatch.\n", tid);
951
952 dispatchInsts(tid);
953 } else if (dispatchStatus[tid] == Unblocking) {
954 // Make sure that the skid buffer has something in it if the
955 // status is unblocking.
956 assert(!skidsEmpty());
957
958 // If the status was unblocking, then instructions from the skid
959 // buffer were used. Remove those instructions and handle
960 // the rest of unblocking.
961 dispatchInsts(tid);
962
963 ++iewUnblockCycles;
964
965 if (validInstsFromRename()) {
966 // Add the current inputs to the skid buffer so they can be
967 // reprocessed when this stage unblocks.
968 skidInsert(tid);
969 }
970
971 unblock(tid);
972 }
973}
974
975template <class Impl>
976void
977DefaultIEW<Impl>::dispatchInsts(ThreadID tid)
978{
979 // Obtain instructions from skid buffer if unblocking, or queue from rename
980 // otherwise.
981 std::queue<DynInstPtr> &insts_to_dispatch =
982 dispatchStatus[tid] == Unblocking ?
983 skidBuffer[tid] : insts[tid];
984
985 int insts_to_add = insts_to_dispatch.size();
986
987 DynInstPtr inst;
988 bool add_to_iq = false;
989 int dis_num_inst = 0;
990
991 // Loop through the instructions, putting them in the instruction
992 // queue.
993 for ( ; dis_num_inst < insts_to_add &&
994 dis_num_inst < dispatchWidth;
995 ++dis_num_inst)
996 {
997 inst = insts_to_dispatch.front();
998
999 if (dispatchStatus[tid] == Unblocking) {
1000 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
1001 "buffer\n", tid);
1002 }
1003
1004 // Make sure there's a valid instruction there.
1005 assert(inst);
1006
1007 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %s [sn:%lli] [tid:%i] to "
1008 "IQ.\n",
1009 tid, inst->pcState(), inst->seqNum, inst->threadNumber);
1010
1011 // Be sure to mark these instructions as ready so that the
1012 // commit stage can go ahead and execute them, and mark
1013 // them as issued so the IQ doesn't reprocess them.
1014
1015 // Check for squashed instructions.
1016 if (inst->isSquashed()) {
1017 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
1018 "not adding to IQ.\n", tid);
1019
1020 ++iewDispSquashedInsts;
1021
1022 insts_to_dispatch.pop();
1023
1024 //Tell Rename That An Instruction has been processed
1025 if (inst->isLoad()) {
1026 toRename->iewInfo[tid].dispatchedToLQ++;
1027 }
1028 if (inst->isStore()) {
1029 toRename->iewInfo[tid].dispatchedToSQ++;
1030 }
1031
1032 toRename->iewInfo[tid].dispatched++;
1033
1034 continue;
1035 }
1036
1037 // Check for full conditions.
1038 if (instQueue.isFull(tid)) {
1039 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1040
1041 // Call function to start blocking.
1042 block(tid);
1043
1044 // Set unblock to false. Special case where we are using
1045 // skidbuffer (unblocking) instructions but then we still
1046 // get full in the IQ.
1047 toRename->iewUnblock[tid] = false;
1048
1049 ++iewIQFullEvents;
1050 break;
1051 }
1052
1053 // Check LSQ if inst is LD/ST
1054 if ((inst->isLoad() && ldstQueue.lqFull(tid)) ||
1055 (inst->isStore() && ldstQueue.sqFull(tid))) {
1056 DPRINTF(IEW, "[tid:%i]: Issue: %s has become full.\n",tid,
1057 inst->isLoad() ? "LQ" : "SQ");
1058
1059 // Call function to start blocking.
1060 block(tid);
1061
1062 // Set unblock to false. Special case where we are using
1063 // skidbuffer (unblocking) instructions but then we still
1064 // get full in the IQ.
1065 toRename->iewUnblock[tid] = false;
1066
1067 ++iewLSQFullEvents;
1068 break;
1069 }
1070
1071 // Otherwise issue the instruction just fine.
1072 if (inst->isLoad()) {
1073 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1074 "encountered, adding to LSQ.\n", tid);
1075
1076 // Reserve a spot in the load store queue for this
1077 // memory access.
1078 ldstQueue.insertLoad(inst);
1079
1080 ++iewDispLoadInsts;
1081
1082 add_to_iq = true;
1083
1084 toRename->iewInfo[tid].dispatchedToLQ++;
1085 } else if (inst->isStore()) {
1086 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1087 "encountered, adding to LSQ.\n", tid);
1088
1089 ldstQueue.insertStore(inst);
1090
1091 ++iewDispStoreInsts;
1092
1093 if (inst->isStoreConditional()) {
1094 // Store conditionals need to be set as "canCommit()"
1095 // so that commit can process them when they reach the
1096 // head of commit.
1097 // @todo: This is somewhat specific to Alpha.
1098 inst->setCanCommit();
1099 instQueue.insertNonSpec(inst);
1100 add_to_iq = false;
1101
1102 ++iewDispNonSpecInsts;
1103 } else {
1104 add_to_iq = true;
1105 }
1106
1107 toRename->iewInfo[tid].dispatchedToSQ++;
1108 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1109 // Same as non-speculative stores.
1110 inst->setCanCommit();
1111 instQueue.insertBarrier(inst);
1112 add_to_iq = false;
1113 } else if (inst->isNop()) {
1114 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1115 "skipping.\n", tid);
1116
1117 inst->setIssued();
1118 inst->setExecuted();
1119 inst->setCanCommit();
1120
1121 instQueue.recordProducer(inst);
1122
1123 iewExecutedNop[tid]++;
1124
1125 add_to_iq = false;
1126 } else if (inst->isExecuted()) {
1127 assert(0 && "Instruction shouldn't be executed.\n");
1128 DPRINTF(IEW, "Issue: Executed branch encountered, "
1129 "skipping.\n");
1130
1131 inst->setIssued();
1132 inst->setCanCommit();
1133
1134 instQueue.recordProducer(inst);
1135
1136 add_to_iq = false;
1137 } else {
1138 add_to_iq = true;
1139 }
1140 if (inst->isNonSpeculative()) {
1141 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1142 "encountered, skipping.\n", tid);
1143
1144 // Same as non-speculative stores.
1145 inst->setCanCommit();
1146
1147 // Specifically insert it as nonspeculative.
1148 instQueue.insertNonSpec(inst);
1149
1150 ++iewDispNonSpecInsts;
1151
1152 add_to_iq = false;
1153 }
1154
1155 // If the instruction queue is not full, then add the
1156 // instruction.
1157 if (add_to_iq) {
1158 instQueue.insert(inst);
1159 }
1160
1161 insts_to_dispatch.pop();
1162
1163 toRename->iewInfo[tid].dispatched++;
1164
1165 ++iewDispatchedInsts;
1166
1167#if TRACING_ON
1168 inst->dispatchTick = curTick() - inst->fetchTick;
1169#endif
1170 ppDispatch->notify(inst);
1171 }
1172
1173 if (!insts_to_dispatch.empty()) {
1174 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1175 block(tid);
1176 toRename->iewUnblock[tid] = false;
1177 }
1178
1179 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1180 dispatchStatus[tid] = Running;
1181
1182 updatedQueues = true;
1183 }
1184
1185 dis_num_inst = 0;
1186}
1187
1188template <class Impl>
1189void
1190DefaultIEW<Impl>::printAvailableInsts()
1191{
1192 int inst = 0;
1193
1194 std::cout << "Available Instructions: ";
1195
1196 while (fromIssue->insts[inst]) {
1197
1198 if (inst%3==0) std::cout << "\n\t";
1199
1200 std::cout << "PC: " << fromIssue->insts[inst]->pcState()
1201 << " TN: " << fromIssue->insts[inst]->threadNumber
1202 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1203
1204 inst++;
1205
1206 }
1207
1208 std::cout << "\n";
1209}
1210
1211template <class Impl>
1212void
1213DefaultIEW<Impl>::executeInsts()
1214{
1215 wbNumInst = 0;
1216 wbCycle = 0;
1217
1218 list<ThreadID>::iterator threads = activeThreads->begin();
1219 list<ThreadID>::iterator end = activeThreads->end();
1220
1221 while (threads != end) {
1222 ThreadID tid = *threads++;
1223 fetchRedirect[tid] = false;
1224 }
1225
1226 // Uncomment this if you want to see all available instructions.
1227 // @todo This doesn't actually work anymore, we should fix it.
1228// printAvailableInsts();
1229
1230 // Execute/writeback any instructions that are available.
1231 int insts_to_execute = fromIssue->size;
1232 int inst_num = 0;
1233 for (; inst_num < insts_to_execute;
1234 ++inst_num) {
1235
1236 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1237
1238 DynInstPtr inst = instQueue.getInstToExecute();
1239
1240 DPRINTF(IEW, "Execute: Processing PC %s, [tid:%i] [sn:%i].\n",
1241 inst->pcState(), inst->threadNumber,inst->seqNum);
1242
1243 // Check if the instruction is squashed; if so then skip it
1244 if (inst->isSquashed()) {
1245 DPRINTF(IEW, "Execute: Instruction was squashed. PC: %s, [tid:%i]"
1246 " [sn:%i]\n", inst->pcState(), inst->threadNumber,
1247 inst->seqNum);
1248
1249 // Consider this instruction executed so that commit can go
1250 // ahead and retire the instruction.
1251 inst->setExecuted();
1252
1253 // Not sure if I should set this here or just let commit try to
1254 // commit any squashed instructions. I like the latter a bit more.
1255 inst->setCanCommit();
1256
1257 ++iewExecSquashedInsts;
1258
1259 continue;
1260 }
1261
1262 Fault fault = NoFault;
1263
1264 // Execute instruction.
1265 // Note that if the instruction faults, it will be handled
1266 // at the commit stage.
1267 if (inst->isMemRef()) {
1268 DPRINTF(IEW, "Execute: Calculating address for memory "
1269 "reference.\n");
1270
1271 // Tell the LDSTQ to execute this instruction (if it is a load).
1272 if (inst->isLoad()) {
1273 // Loads will mark themselves as executed, and their writeback
1274 // event adds the instruction to the queue to commit
1275 fault = ldstQueue.executeLoad(inst);
1276
1277 if (inst->isTranslationDelayed() &&
1278 fault == NoFault) {
1279 // A hw page table walk is currently going on; the
1280 // instruction must be deferred.
1281 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1282 "load.\n");
1283 instQueue.deferMemInst(inst);
1284 continue;
1285 }
1286
1287 if (inst->isDataPrefetch() || inst->isInstPrefetch()) {
1288 inst->fault = NoFault;
1289 }
1290 } else if (inst->isStore()) {
1291 fault = ldstQueue.executeStore(inst);
1292
1293 if (inst->isTranslationDelayed() &&
1294 fault == NoFault) {
1295 // A hw page table walk is currently going on; the
1296 // instruction must be deferred.
1297 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1298 "store.\n");
1299 instQueue.deferMemInst(inst);
1300 continue;
1301 }
1302
1303 // If the store had a fault then it may not have a mem req
1304 if (fault != NoFault || !inst->readPredicate() ||
1305 !inst->isStoreConditional()) {
1306 // If the instruction faulted, then we need to send it along
1307 // to commit without the instruction completing.
1308 // Send this instruction to commit, also make sure iew stage
1309 // realizes there is activity.
1310 inst->setExecuted();
1311 instToCommit(inst);
1312 activityThisCycle();
1313 }
1314
1315 // Store conditionals will mark themselves as
1316 // executed, and their writeback event will add the
1317 // instruction to the queue to commit.
1318 } else {
1319 panic("Unexpected memory type!\n");
1320 }
1321
1322 } else {
1323 // If the instruction has already faulted, then skip executing it.
1324 // Such case can happen when it faulted during ITLB translation.
1325 // If we execute the instruction (even if it's a nop) the fault
1326 // will be replaced and we will lose it.
1327 if (inst->getFault() == NoFault) {
1328 inst->execute();
1329 if (!inst->readPredicate())
1330 inst->forwardOldRegs();
1331 }
1332
1333 inst->setExecuted();
1334
1335 instToCommit(inst);
1336 }
1337
1338 updateExeInstStats(inst);
1339
1340 // Check if branch prediction was correct, if not then we need
1341 // to tell commit to squash in flight instructions. Only
1342 // handle this if there hasn't already been something that
1343 // redirects fetch in this group of instructions.
1344
1345 // This probably needs to prioritize the redirects if a different
1346 // scheduler is used. Currently the scheduler schedules the oldest
1347 // instruction first, so the branch resolution order will be correct.
1348 ThreadID tid = inst->threadNumber;
1349
1350 if (!fetchRedirect[tid] ||
1351 !toCommit->squash[tid] ||
1352 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1353
1354 // Prevent testing for misprediction on load instructions,
1355 // that have not been executed.
1356 bool loadNotExecuted = !inst->isExecuted() && inst->isLoad();
1357
1358 if (inst->mispredicted() && !loadNotExecuted) {
1359 fetchRedirect[tid] = true;
1360
1361 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1362 DPRINTF(IEW, "Predicted target was PC: %s.\n",
1363 inst->readPredTarg());
1364 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %s.\n",
1365 inst->pcState());
1366 // If incorrect, then signal the ROB that it must be squashed.
1367 squashDueToBranch(inst, tid);
1368
1369 ppMispredict->notify(inst);
1370
1371 if (inst->readPredTaken()) {
1372 predictedTakenIncorrect++;
1373 } else {
1374 predictedNotTakenIncorrect++;
1375 }
1376 } else if (ldstQueue.violation(tid)) {
1377 assert(inst->isMemRef());
1378 // If there was an ordering violation, then get the
1379 // DynInst that caused the violation. Note that this
1380 // clears the violation signal.
1381 DynInstPtr violator;
1382 violator = ldstQueue.getMemDepViolator(tid);
1383
1384 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: %s "
1385 "[sn:%lli], inst PC: %s [sn:%lli]. Addr is: %#x.\n",
1386 violator->pcState(), violator->seqNum,
1387 inst->pcState(), inst->seqNum, inst->physEffAddr);
1388
1389 fetchRedirect[tid] = true;
1390
1391 // Tell the instruction queue that a violation has occured.
1392 instQueue.violation(inst, violator);
1393
1394 // Squash.
1395 squashDueToMemOrder(violator, tid);
1396
1397 ++memOrderViolationEvents;
1398 } else if (ldstQueue.loadBlocked(tid) &&
1399 !ldstQueue.isLoadBlockedHandled(tid)) {
1400 fetchRedirect[tid] = true;
1401
1402 DPRINTF(IEW, "Load operation couldn't execute because the "
1403 "memory system is blocked. PC: %s [sn:%lli]\n",
1404 inst->pcState(), inst->seqNum);
1405
1406 squashDueToMemBlocked(inst, tid);
1407 }
1408 } else {
1409 // Reset any state associated with redirects that will not
1410 // be used.
1411 if (ldstQueue.violation(tid)) {
1412 assert(inst->isMemRef());
1413
1414 DynInstPtr violator = ldstQueue.getMemDepViolator(tid);
1415
1416 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1417 "%s, inst PC: %s. Addr is: %#x.\n",
1418 violator->pcState(), inst->pcState(),
1419 inst->physEffAddr);
1420 DPRINTF(IEW, "Violation will not be handled because "
1421 "already squashing\n");
1422
1423 ++memOrderViolationEvents;
1424 }
1425 if (ldstQueue.loadBlocked(tid) &&
1426 !ldstQueue.isLoadBlockedHandled(tid)) {
1427 DPRINTF(IEW, "Load operation couldn't execute because the "
1428 "memory system is blocked. PC: %s [sn:%lli]\n",
1429 inst->pcState(), inst->seqNum);
1430 DPRINTF(IEW, "Blocked load will not be handled because "
1431 "already squashing\n");
1432
1433 ldstQueue.setLoadBlockedHandled(tid);
1434 }
1435
1436 }
1437 }
1438
1439 // Update and record activity if we processed any instructions.
1440 if (inst_num) {
1441 if (exeStatus == Idle) {
1442 exeStatus = Running;
1443 }
1444
1445 updatedQueues = true;
1446
1447 cpu->activityThisCycle();
1448 }
1449
1450 // Need to reset this in case a writeback event needs to write into the
1451 // iew queue. That way the writeback event will write into the correct
1452 // spot in the queue.
1453 wbNumInst = 0;
1454
1455}
1456
1457template <class Impl>
1458void
1459DefaultIEW<Impl>::writebackInsts()
1460{
1461 // Loop through the head of the time buffer and wake any
1462 // dependents. These instructions are about to write back. Also
1463 // mark scoreboard that this instruction is finally complete.
1464 // Either have IEW have direct access to scoreboard, or have this
1465 // as part of backwards communication.
1466 for (int inst_num = 0; inst_num < wbWidth &&
1467 toCommit->insts[inst_num]; inst_num++) {
1468 DynInstPtr inst = toCommit->insts[inst_num];
1469 ThreadID tid = inst->threadNumber;
1470
1471 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %s.\n",
1472 inst->seqNum, inst->pcState());
1473
1474 iewInstsToCommit[tid]++;
1475
1476 // Some instructions will be sent to commit without having
1477 // executed because they need commit to handle them.
1478 // E.g. Uncached loads have not actually executed when they
1479 // are first sent to commit. Instead commit must tell the LSQ
1480 // when it's ready to execute the uncached load.
1481 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1482 int dependents = instQueue.wakeDependents(inst);
1483
1484 for (int i = 0; i < inst->numDestRegs(); i++) {
1485 //mark as Ready
1486 DPRINTF(IEW,"Setting Destination Register %i\n",
1487 inst->renamedDestRegIdx(i));
1488 scoreboard->setReg(inst->renamedDestRegIdx(i));
1489 }
1490
1491 if (dependents) {
1492 producerInst[tid]++;
1493 consumerInst[tid]+= dependents;
1494 }
1495 writebackCount[tid]++;
1496 }
1497 }
1498}
1499
1500template<class Impl>
1501void
1502DefaultIEW<Impl>::tick()
1503{
1504 wbNumInst = 0;
1505 wbCycle = 0;
1506
1507 wroteToTimeBuffer = false;
1508 updatedQueues = false;
1509
1510 sortInsts();
1511
1512 // Free function units marked as being freed this cycle.
1513 fuPool->processFreeUnits();
1514
1515 list<ThreadID>::iterator threads = activeThreads->begin();
1516 list<ThreadID>::iterator end = activeThreads->end();
1517
1518 // Check stall and squash signals, dispatch any instructions.
1519 while (threads != end) {
1520 ThreadID tid = *threads++;
1521
1522 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1523
1524 checkSignalsAndUpdate(tid);
1525 dispatch(tid);
1526 }
1527
1528 if (exeStatus != Squashing) {
1529 executeInsts();
1530
1531 writebackInsts();
1532
1533 // Have the instruction queue try to schedule any ready instructions.
1534 // (In actuality, this scheduling is for instructions that will
1535 // be executed next cycle.)
1536 instQueue.scheduleReadyInsts();
1537
1538 // Also should advance its own time buffers if the stage ran.
1539 // Not the best place for it, but this works (hopefully).
1540 issueToExecQueue.advance();
1541 }
1542
1543 bool broadcast_free_entries = false;
1544
1545 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1546 exeStatus = Idle;
1547 updateLSQNextCycle = false;
1548
1549 broadcast_free_entries = true;
1550 }
1551
1552 // Writeback any stores using any leftover bandwidth.
1553 ldstQueue.writebackStores();
1554
1555 // Check the committed load/store signals to see if there's a load
1556 // or store to commit. Also check if it's being told to execute a
1557 // nonspeculative instruction.
1558 // This is pretty inefficient...
1559
1560 threads = activeThreads->begin();
1561 while (threads != end) {
1562 ThreadID tid = (*threads++);
1563
1564 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1565
1566 // Update structures based on instructions committed.
1567 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1568 !fromCommit->commitInfo[tid].squash &&
1569 !fromCommit->commitInfo[tid].robSquashing) {
1570
1571 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1572
1573 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1574
1575 updateLSQNextCycle = true;
1576 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1577 }
1578
1579 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1580
1581 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1582 if (fromCommit->commitInfo[tid].uncached) {
1583 instQueue.replayMemInst(fromCommit->commitInfo[tid].uncachedLoad);
1584 fromCommit->commitInfo[tid].uncachedLoad->setAtCommit();
1585 } else {
1586 instQueue.scheduleNonSpec(
1587 fromCommit->commitInfo[tid].nonSpecSeqNum);
1588 }
1589 }
1590
1591 if (broadcast_free_entries) {
1592 toFetch->iewInfo[tid].iqCount =
1593 instQueue.getCount(tid);
1594 toFetch->iewInfo[tid].ldstqCount =
1595 ldstQueue.getCount(tid);
1596
1597 toRename->iewInfo[tid].usedIQ = true;
1598 toRename->iewInfo[tid].freeIQEntries =
1599 instQueue.numFreeEntries(tid);
1600 toRename->iewInfo[tid].usedLSQ = true;
1601
1602 toRename->iewInfo[tid].freeLQEntries =
1603 ldstQueue.numFreeLoadEntries(tid);
1604 toRename->iewInfo[tid].freeSQEntries =
1605 ldstQueue.numFreeStoreEntries(tid);
1606
1607 wroteToTimeBuffer = true;
1608 }
1609
1610 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1611 tid, toRename->iewInfo[tid].dispatched);
1612 }
1613
1614 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1615 "LQ has %i free entries. SQ has %i free entries.\n",
1616 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1617 ldstQueue.numFreeLoadEntries(), ldstQueue.numFreeStoreEntries());
1618
1619 updateStatus();
1620
1621 if (wroteToTimeBuffer) {
1622 DPRINTF(Activity, "Activity this cycle.\n");
1623 cpu->activityThisCycle();
1624 }
1625}
1626
1627template <class Impl>
1628void
1629DefaultIEW<Impl>::updateExeInstStats(DynInstPtr &inst)
1630{
1631 ThreadID tid = inst->threadNumber;
1632
1633 iewExecutedInsts++;
1634
1635#if TRACING_ON
1636 if (DTRACE(O3PipeView)) {
1637 inst->completeTick = curTick() - inst->fetchTick;
1638 }
1639#endif
1640
1641 //
1642 // Control operations
1643 //
1644 if (inst->isControl())
1645 iewExecutedBranches[tid]++;
1646
1647 //
1648 // Memory operations
1649 //
1650 if (inst->isMemRef()) {
1651 iewExecutedRefs[tid]++;
1652
1653 if (inst->isLoad()) {
1654 iewExecLoadInsts[tid]++;
1655 }
1656 }
1657}
1658
1659template <class Impl>
1660void
1661DefaultIEW<Impl>::checkMisprediction(DynInstPtr &inst)
1662{
1663 ThreadID tid = inst->threadNumber;
1664
1665 if (!fetchRedirect[tid] ||
1666 !toCommit->squash[tid] ||
1667 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1668
1669 if (inst->mispredicted()) {
1670 fetchRedirect[tid] = true;
1671
1672 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1673 DPRINTF(IEW, "Predicted target was PC:%#x, NPC:%#x.\n",
1674 inst->predInstAddr(), inst->predNextInstAddr());
1675 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x,"
1676 " NPC: %#x.\n", inst->nextInstAddr(),
1677 inst->nextInstAddr());
1678 // If incorrect, then signal the ROB that it must be squashed.
1679 squashDueToBranch(inst, tid);
1680
1681 if (inst->readPredTaken()) {
1682 predictedTakenIncorrect++;
1683 } else {
1684 predictedNotTakenIncorrect++;
1685 }
1686 }
1687 }
1688}
1689
1690#endif//__CPU_O3_IEW_IMPL_IMPL_HH__