cpu: fixed how O3 CPU executes an exit system call
[gem5.git] / src / cpu / o3 / iew_impl.hh
1 /*
2 * Copyright (c) 2010-2013, 2018 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
65 using namespace std;
66
67 template<class Impl>
68 DefaultIEW<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 wbNumInst(0),
80 wbCycle(0),
81 wbWidth(params->wbWidth),
82 numThreads(params->numThreads)
83 {
84 if (dispatchWidth > Impl::MaxWidth)
85 fatal("dispatchWidth (%d) is larger than compiled limit (%d),\n"
86 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
87 dispatchWidth, static_cast<int>(Impl::MaxWidth));
88 if (issueWidth > Impl::MaxWidth)
89 fatal("issueWidth (%d) is larger than compiled limit (%d),\n"
90 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
91 issueWidth, static_cast<int>(Impl::MaxWidth));
92 if (wbWidth > Impl::MaxWidth)
93 fatal("wbWidth (%d) is larger than compiled limit (%d),\n"
94 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
95 wbWidth, static_cast<int>(Impl::MaxWidth));
96
97 _status = Active;
98 exeStatus = Running;
99 wbStatus = Idle;
100
101 // Setup wire to read instructions coming from issue.
102 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
103
104 // Instruction queue needs the queue between issue and execute.
105 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
106
107 for (ThreadID tid = 0; tid < Impl::MaxThreads; tid++) {
108 dispatchStatus[tid] = Running;
109 fetchRedirect[tid] = false;
110 }
111
112 updateLSQNextCycle = false;
113
114 skidBufferMax = (renameToIEWDelay + 1) * params->renameWidth;
115 }
116
117 template <class Impl>
118 std::string
119 DefaultIEW<Impl>::name() const
120 {
121 return cpu->name() + ".iew";
122 }
123
124 template <class Impl>
125 void
126 DefaultIEW<Impl>::regProbePoints()
127 {
128 ppDispatch = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Dispatch");
129 ppMispredict = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Mispredict");
130 /**
131 * Probe point with dynamic instruction as the argument used to probe when
132 * an instruction starts to execute.
133 */
134 ppExecute = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(),
135 "Execute");
136 /**
137 * Probe point with dynamic instruction as the argument used to probe when
138 * an instruction execution completes and it is marked ready to commit.
139 */
140 ppToCommit = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(),
141 "ToCommit");
142 }
143
144 template <class Impl>
145 void
146 DefaultIEW<Impl>::regStats()
147 {
148 using namespace Stats;
149
150 instQueue.regStats();
151 ldstQueue.regStats();
152
153 iewIdleCycles
154 .name(name() + ".iewIdleCycles")
155 .desc("Number of cycles IEW is idle");
156
157 iewSquashCycles
158 .name(name() + ".iewSquashCycles")
159 .desc("Number of cycles IEW is squashing");
160
161 iewBlockCycles
162 .name(name() + ".iewBlockCycles")
163 .desc("Number of cycles IEW is blocking");
164
165 iewUnblockCycles
166 .name(name() + ".iewUnblockCycles")
167 .desc("Number of cycles IEW is unblocking");
168
169 iewDispatchedInsts
170 .name(name() + ".iewDispatchedInsts")
171 .desc("Number of instructions dispatched to IQ");
172
173 iewDispSquashedInsts
174 .name(name() + ".iewDispSquashedInsts")
175 .desc("Number of squashed instructions skipped by dispatch");
176
177 iewDispLoadInsts
178 .name(name() + ".iewDispLoadInsts")
179 .desc("Number of dispatched load instructions");
180
181 iewDispStoreInsts
182 .name(name() + ".iewDispStoreInsts")
183 .desc("Number of dispatched store instructions");
184
185 iewDispNonSpecInsts
186 .name(name() + ".iewDispNonSpecInsts")
187 .desc("Number of dispatched non-speculative instructions");
188
189 iewIQFullEvents
190 .name(name() + ".iewIQFullEvents")
191 .desc("Number of times the IQ has become full, causing a stall");
192
193 iewLSQFullEvents
194 .name(name() + ".iewLSQFullEvents")
195 .desc("Number of times the LSQ has become full, causing a stall");
196
197 memOrderViolationEvents
198 .name(name() + ".memOrderViolationEvents")
199 .desc("Number of memory order violations");
200
201 predictedTakenIncorrect
202 .name(name() + ".predictedTakenIncorrect")
203 .desc("Number of branches that were predicted taken incorrectly");
204
205 predictedNotTakenIncorrect
206 .name(name() + ".predictedNotTakenIncorrect")
207 .desc("Number of branches that were predicted not taken incorrectly");
208
209 branchMispredicts
210 .name(name() + ".branchMispredicts")
211 .desc("Number of branch mispredicts detected at execute");
212
213 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
214
215 iewExecutedInsts
216 .name(name() + ".iewExecutedInsts")
217 .desc("Number of executed instructions");
218
219 iewExecLoadInsts
220 .init(cpu->numThreads)
221 .name(name() + ".iewExecLoadInsts")
222 .desc("Number of load instructions executed")
223 .flags(total);
224
225 iewExecSquashedInsts
226 .name(name() + ".iewExecSquashedInsts")
227 .desc("Number of squashed instructions skipped in execute");
228
229 iewExecutedSwp
230 .init(cpu->numThreads)
231 .name(name() + ".exec_swp")
232 .desc("number of swp insts executed")
233 .flags(total);
234
235 iewExecutedNop
236 .init(cpu->numThreads)
237 .name(name() + ".exec_nop")
238 .desc("number of nop insts executed")
239 .flags(total);
240
241 iewExecutedRefs
242 .init(cpu->numThreads)
243 .name(name() + ".exec_refs")
244 .desc("number of memory reference insts executed")
245 .flags(total);
246
247 iewExecutedBranches
248 .init(cpu->numThreads)
249 .name(name() + ".exec_branches")
250 .desc("Number of branches executed")
251 .flags(total);
252
253 iewExecStoreInsts
254 .name(name() + ".exec_stores")
255 .desc("Number of stores executed")
256 .flags(total);
257 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
258
259 iewExecRate
260 .name(name() + ".exec_rate")
261 .desc("Inst execution rate")
262 .flags(total);
263
264 iewExecRate = iewExecutedInsts / cpu->numCycles;
265
266 iewInstsToCommit
267 .init(cpu->numThreads)
268 .name(name() + ".wb_sent")
269 .desc("cumulative count of insts sent to commit")
270 .flags(total);
271
272 writebackCount
273 .init(cpu->numThreads)
274 .name(name() + ".wb_count")
275 .desc("cumulative count of insts written-back")
276 .flags(total);
277
278 producerInst
279 .init(cpu->numThreads)
280 .name(name() + ".wb_producers")
281 .desc("num instructions producing a value")
282 .flags(total);
283
284 consumerInst
285 .init(cpu->numThreads)
286 .name(name() + ".wb_consumers")
287 .desc("num instructions consuming a value")
288 .flags(total);
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
304 template<class Impl>
305 void
306 DefaultIEW<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
326 template<class Impl>
327 void
328 DefaultIEW<Impl>::clearStates(ThreadID tid)
329 {
330 toRename->iewInfo[tid].usedIQ = true;
331 toRename->iewInfo[tid].freeIQEntries =
332 instQueue.numFreeEntries(tid);
333
334 toRename->iewInfo[tid].usedLSQ = true;
335 toRename->iewInfo[tid].freeLQEntries = ldstQueue.numFreeLoadEntries(tid);
336 toRename->iewInfo[tid].freeSQEntries = ldstQueue.numFreeStoreEntries(tid);
337 }
338
339 template<class Impl>
340 void
341 DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
342 {
343 timeBuffer = tb_ptr;
344
345 // Setup wire to read information from time buffer, from commit.
346 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
347
348 // Setup wire to write information back to previous stages.
349 toRename = timeBuffer->getWire(0);
350
351 toFetch = timeBuffer->getWire(0);
352
353 // Instruction queue also needs main time buffer.
354 instQueue.setTimeBuffer(tb_ptr);
355 }
356
357 template<class Impl>
358 void
359 DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
360 {
361 renameQueue = rq_ptr;
362
363 // Setup wire to read information from rename queue.
364 fromRename = renameQueue->getWire(-renameToIEWDelay);
365 }
366
367 template<class Impl>
368 void
369 DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
370 {
371 iewQueue = iq_ptr;
372
373 // Setup wire to write instructions to commit.
374 toCommit = iewQueue->getWire(0);
375 }
376
377 template<class Impl>
378 void
379 DefaultIEW<Impl>::setActiveThreads(list<ThreadID> *at_ptr)
380 {
381 activeThreads = at_ptr;
382
383 ldstQueue.setActiveThreads(at_ptr);
384 instQueue.setActiveThreads(at_ptr);
385 }
386
387 template<class Impl>
388 void
389 DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
390 {
391 scoreboard = sb_ptr;
392 }
393
394 template <class Impl>
395 bool
396 DefaultIEW<Impl>::isDrained() const
397 {
398 bool drained = ldstQueue.isDrained() && instQueue.isDrained();
399
400 for (ThreadID tid = 0; tid < numThreads; tid++) {
401 if (!insts[tid].empty()) {
402 DPRINTF(Drain, "%i: Insts not empty.\n", tid);
403 drained = false;
404 }
405 if (!skidBuffer[tid].empty()) {
406 DPRINTF(Drain, "%i: Skid buffer not empty.\n", tid);
407 drained = false;
408 }
409 drained = drained && dispatchStatus[tid] == Running;
410 }
411
412 // Also check the FU pool as instructions are "stored" in FU
413 // completion events until they are done and not accounted for
414 // above
415 if (drained && !fuPool->isDrained()) {
416 DPRINTF(Drain, "FU pool still busy.\n");
417 drained = false;
418 }
419
420 return drained;
421 }
422
423 template <class Impl>
424 void
425 DefaultIEW<Impl>::drainSanityCheck() const
426 {
427 assert(isDrained());
428
429 instQueue.drainSanityCheck();
430 ldstQueue.drainSanityCheck();
431 }
432
433 template <class Impl>
434 void
435 DefaultIEW<Impl>::takeOverFrom()
436 {
437 // Reset all state.
438 _status = Active;
439 exeStatus = Running;
440 wbStatus = Idle;
441
442 instQueue.takeOverFrom();
443 ldstQueue.takeOverFrom();
444 fuPool->takeOverFrom();
445
446 startupStage();
447 cpu->activityThisCycle();
448
449 for (ThreadID tid = 0; tid < numThreads; tid++) {
450 dispatchStatus[tid] = Running;
451 fetchRedirect[tid] = false;
452 }
453
454 updateLSQNextCycle = false;
455
456 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
457 issueToExecQueue.advance();
458 }
459 }
460
461 template<class Impl>
462 void
463 DefaultIEW<Impl>::squash(ThreadID tid)
464 {
465 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n", tid);
466
467 // Tell the IQ to start squashing.
468 instQueue.squash(tid);
469
470 // Tell the LDSTQ to start squashing.
471 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
472 updatedQueues = true;
473
474 // Clear the skid buffer in case it has any data in it.
475 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
476 tid, fromCommit->commitInfo[tid].doneSeqNum);
477
478 while (!skidBuffer[tid].empty()) {
479 if (skidBuffer[tid].front()->isLoad()) {
480 toRename->iewInfo[tid].dispatchedToLQ++;
481 }
482 if (skidBuffer[tid].front()->isStore()) {
483 toRename->iewInfo[tid].dispatchedToSQ++;
484 }
485
486 toRename->iewInfo[tid].dispatched++;
487
488 skidBuffer[tid].pop();
489 }
490
491 emptyRenameInsts(tid);
492 }
493
494 template<class Impl>
495 void
496 DefaultIEW<Impl>::squashDueToBranch(const DynInstPtr& inst, ThreadID tid)
497 {
498 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %s "
499 "[sn:%i].\n", tid, inst->pcState(), inst->seqNum);
500
501 if (!toCommit->squash[tid] ||
502 inst->seqNum < toCommit->squashedSeqNum[tid]) {
503 toCommit->squash[tid] = true;
504 toCommit->squashedSeqNum[tid] = inst->seqNum;
505 toCommit->branchTaken[tid] = inst->pcState().branching();
506
507 TheISA::PCState pc = inst->pcState();
508 TheISA::advancePC(pc, inst->staticInst);
509
510 toCommit->pc[tid] = pc;
511 toCommit->mispredictInst[tid] = inst;
512 toCommit->includeSquashInst[tid] = false;
513
514 wroteToTimeBuffer = true;
515 }
516
517 }
518
519 template<class Impl>
520 void
521 DefaultIEW<Impl>::squashDueToMemOrder(const DynInstPtr& inst, ThreadID tid)
522 {
523 DPRINTF(IEW, "[tid:%i]: Memory violation, squashing violator and younger "
524 "insts, PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
525 // Need to include inst->seqNum in the following comparison to cover the
526 // corner case when a branch misprediction and a memory violation for the
527 // same instruction (e.g. load PC) are detected in the same cycle. In this
528 // case the memory violator should take precedence over the branch
529 // misprediction because it requires the violator itself to be included in
530 // the squash.
531 if (!toCommit->squash[tid] ||
532 inst->seqNum <= toCommit->squashedSeqNum[tid]) {
533 toCommit->squash[tid] = true;
534
535 toCommit->squashedSeqNum[tid] = inst->seqNum;
536 toCommit->pc[tid] = inst->pcState();
537 toCommit->mispredictInst[tid] = NULL;
538
539 // Must include the memory violator in the squash.
540 toCommit->includeSquashInst[tid] = true;
541
542 wroteToTimeBuffer = true;
543 }
544 }
545
546 template<class Impl>
547 void
548 DefaultIEW<Impl>::block(ThreadID tid)
549 {
550 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
551
552 if (dispatchStatus[tid] != Blocked &&
553 dispatchStatus[tid] != Unblocking) {
554 toRename->iewBlock[tid] = true;
555 wroteToTimeBuffer = true;
556 }
557
558 // Add the current inputs to the skid buffer so they can be
559 // reprocessed when this stage unblocks.
560 skidInsert(tid);
561
562 dispatchStatus[tid] = Blocked;
563 }
564
565 template<class Impl>
566 void
567 DefaultIEW<Impl>::unblock(ThreadID tid)
568 {
569 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
570 "buffer %u.\n",tid, tid);
571
572 // If the skid bufffer is empty, signal back to previous stages to unblock.
573 // Also switch status to running.
574 if (skidBuffer[tid].empty()) {
575 toRename->iewUnblock[tid] = true;
576 wroteToTimeBuffer = true;
577 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
578 dispatchStatus[tid] = Running;
579 }
580 }
581
582 template<class Impl>
583 void
584 DefaultIEW<Impl>::wakeDependents(const DynInstPtr& inst)
585 {
586 instQueue.wakeDependents(inst);
587 }
588
589 template<class Impl>
590 void
591 DefaultIEW<Impl>::rescheduleMemInst(const DynInstPtr& inst)
592 {
593 instQueue.rescheduleMemInst(inst);
594 }
595
596 template<class Impl>
597 void
598 DefaultIEW<Impl>::replayMemInst(const DynInstPtr& inst)
599 {
600 instQueue.replayMemInst(inst);
601 }
602
603 template<class Impl>
604 void
605 DefaultIEW<Impl>::blockMemInst(const DynInstPtr& inst)
606 {
607 instQueue.blockMemInst(inst);
608 }
609
610 template<class Impl>
611 void
612 DefaultIEW<Impl>::cacheUnblocked()
613 {
614 instQueue.cacheUnblocked();
615 }
616
617 template<class Impl>
618 void
619 DefaultIEW<Impl>::instToCommit(const DynInstPtr& inst)
620 {
621 // This function should not be called after writebackInsts in a
622 // single cycle. That will cause problems with an instruction
623 // being added to the queue to commit without being processed by
624 // writebackInsts prior to being sent to commit.
625
626 // First check the time slot that this instruction will write
627 // to. If there are free write ports at the time, then go ahead
628 // and write the instruction to that time. If there are not,
629 // keep looking back to see where's the first time there's a
630 // free slot.
631 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
632 ++wbNumInst;
633 if (wbNumInst == wbWidth) {
634 ++wbCycle;
635 wbNumInst = 0;
636 }
637 }
638
639 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
640 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
641 // Add finished instruction to queue to commit.
642 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
643 (*iewQueue)[wbCycle].size++;
644 }
645
646 template <class Impl>
647 unsigned
648 DefaultIEW<Impl>::validInstsFromRename()
649 {
650 unsigned inst_count = 0;
651
652 for (int i=0; i<fromRename->size; i++) {
653 if (!fromRename->insts[i]->isSquashed())
654 inst_count++;
655 }
656
657 return inst_count;
658 }
659
660 template<class Impl>
661 void
662 DefaultIEW<Impl>::skidInsert(ThreadID tid)
663 {
664 DynInstPtr inst = NULL;
665
666 while (!insts[tid].empty()) {
667 inst = insts[tid].front();
668
669 insts[tid].pop();
670
671 DPRINTF(IEW,"[tid:%i]: Inserting [sn:%lli] PC:%s into "
672 "dispatch skidBuffer %i\n",tid, inst->seqNum,
673 inst->pcState(),tid);
674
675 skidBuffer[tid].push(inst);
676 }
677
678 assert(skidBuffer[tid].size() <= skidBufferMax &&
679 "Skidbuffer Exceeded Max Size");
680 }
681
682 template<class Impl>
683 int
684 DefaultIEW<Impl>::skidCount()
685 {
686 int max=0;
687
688 list<ThreadID>::iterator threads = activeThreads->begin();
689 list<ThreadID>::iterator end = activeThreads->end();
690
691 while (threads != end) {
692 ThreadID tid = *threads++;
693 unsigned thread_count = skidBuffer[tid].size();
694 if (max < thread_count)
695 max = thread_count;
696 }
697
698 return max;
699 }
700
701 template<class Impl>
702 bool
703 DefaultIEW<Impl>::skidsEmpty()
704 {
705 list<ThreadID>::iterator threads = activeThreads->begin();
706 list<ThreadID>::iterator end = activeThreads->end();
707
708 while (threads != end) {
709 ThreadID tid = *threads++;
710
711 if (!skidBuffer[tid].empty())
712 return false;
713 }
714
715 return true;
716 }
717
718 template <class Impl>
719 void
720 DefaultIEW<Impl>::updateStatus()
721 {
722 bool any_unblocking = false;
723
724 list<ThreadID>::iterator threads = activeThreads->begin();
725 list<ThreadID>::iterator end = activeThreads->end();
726
727 while (threads != end) {
728 ThreadID tid = *threads++;
729
730 if (dispatchStatus[tid] == Unblocking) {
731 any_unblocking = true;
732 break;
733 }
734 }
735
736 // If there are no ready instructions waiting to be scheduled by the IQ,
737 // and there's no stores waiting to write back, and dispatch is not
738 // unblocking, then there is no internal activity for the IEW stage.
739 instQueue.intInstQueueReads++;
740 if (_status == Active && !instQueue.hasReadyInsts() &&
741 !ldstQueue.willWB() && !any_unblocking) {
742 DPRINTF(IEW, "IEW switching to idle\n");
743
744 deactivateStage();
745
746 _status = Inactive;
747 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
748 ldstQueue.willWB() ||
749 any_unblocking)) {
750 // Otherwise there is internal activity. Set to active.
751 DPRINTF(IEW, "IEW switching to active\n");
752
753 activateStage();
754
755 _status = Active;
756 }
757 }
758
759 template <class Impl>
760 bool
761 DefaultIEW<Impl>::checkStall(ThreadID tid)
762 {
763 bool ret_val(false);
764
765 if (fromCommit->commitInfo[tid].robSquashing) {
766 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
767 ret_val = true;
768 } else if (instQueue.isFull(tid)) {
769 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
770 ret_val = true;
771 }
772
773 return ret_val;
774 }
775
776 template <class Impl>
777 void
778 DefaultIEW<Impl>::checkSignalsAndUpdate(ThreadID tid)
779 {
780 // Check if there's a squash signal, squash if there is
781 // Check stall signals, block if there is.
782 // If status was Blocked
783 // if so then go to unblocking
784 // If status was Squashing
785 // check if squashing is not high. Switch to running this cycle.
786
787 if (fromCommit->commitInfo[tid].squash) {
788 squash(tid);
789
790 if (dispatchStatus[tid] == Blocked ||
791 dispatchStatus[tid] == Unblocking) {
792 toRename->iewUnblock[tid] = true;
793 wroteToTimeBuffer = true;
794 }
795
796 dispatchStatus[tid] = Squashing;
797 fetchRedirect[tid] = false;
798 return;
799 }
800
801 if (fromCommit->commitInfo[tid].robSquashing) {
802 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
803
804 dispatchStatus[tid] = Squashing;
805 emptyRenameInsts(tid);
806 wroteToTimeBuffer = true;
807 }
808
809 if (checkStall(tid)) {
810 block(tid);
811 dispatchStatus[tid] = Blocked;
812 return;
813 }
814
815 if (dispatchStatus[tid] == Blocked) {
816 // Status from previous cycle was blocked, but there are no more stall
817 // conditions. Switch over to unblocking.
818 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
819 tid);
820
821 dispatchStatus[tid] = Unblocking;
822
823 unblock(tid);
824
825 return;
826 }
827
828 if (dispatchStatus[tid] == Squashing) {
829 // Switch status to running if rename isn't being told to block or
830 // squash this cycle.
831 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
832 tid);
833
834 dispatchStatus[tid] = Running;
835
836 return;
837 }
838 }
839
840 template <class Impl>
841 void
842 DefaultIEW<Impl>::sortInsts()
843 {
844 int insts_from_rename = fromRename->size;
845 #ifdef DEBUG
846 for (ThreadID tid = 0; tid < numThreads; tid++)
847 assert(insts[tid].empty());
848 #endif
849 for (int i = 0; i < insts_from_rename; ++i) {
850 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
851 }
852 }
853
854 template <class Impl>
855 void
856 DefaultIEW<Impl>::emptyRenameInsts(ThreadID tid)
857 {
858 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions\n", tid);
859
860 while (!insts[tid].empty()) {
861
862 if (insts[tid].front()->isLoad()) {
863 toRename->iewInfo[tid].dispatchedToLQ++;
864 }
865 if (insts[tid].front()->isStore()) {
866 toRename->iewInfo[tid].dispatchedToSQ++;
867 }
868
869 toRename->iewInfo[tid].dispatched++;
870
871 insts[tid].pop();
872 }
873 }
874
875 template <class Impl>
876 void
877 DefaultIEW<Impl>::wakeCPU()
878 {
879 cpu->wakeCPU();
880 }
881
882 template <class Impl>
883 void
884 DefaultIEW<Impl>::activityThisCycle()
885 {
886 DPRINTF(Activity, "Activity this cycle.\n");
887 cpu->activityThisCycle();
888 }
889
890 template <class Impl>
891 inline void
892 DefaultIEW<Impl>::activateStage()
893 {
894 DPRINTF(Activity, "Activating stage.\n");
895 cpu->activateStage(O3CPU::IEWIdx);
896 }
897
898 template <class Impl>
899 inline void
900 DefaultIEW<Impl>::deactivateStage()
901 {
902 DPRINTF(Activity, "Deactivating stage.\n");
903 cpu->deactivateStage(O3CPU::IEWIdx);
904 }
905
906 template<class Impl>
907 void
908 DefaultIEW<Impl>::dispatch(ThreadID tid)
909 {
910 // If status is Running or idle,
911 // call dispatchInsts()
912 // If status is Unblocking,
913 // buffer any instructions coming from rename
914 // continue trying to empty skid buffer
915 // check if stall conditions have passed
916
917 if (dispatchStatus[tid] == Blocked) {
918 ++iewBlockCycles;
919
920 } else if (dispatchStatus[tid] == Squashing) {
921 ++iewSquashCycles;
922 }
923
924 // Dispatch should try to dispatch as many instructions as its bandwidth
925 // will allow, as long as it is not currently blocked.
926 if (dispatchStatus[tid] == Running ||
927 dispatchStatus[tid] == Idle) {
928 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
929 "dispatch.\n", tid);
930
931 dispatchInsts(tid);
932 } else if (dispatchStatus[tid] == Unblocking) {
933 // Make sure that the skid buffer has something in it if the
934 // status is unblocking.
935 assert(!skidsEmpty());
936
937 // If the status was unblocking, then instructions from the skid
938 // buffer were used. Remove those instructions and handle
939 // the rest of unblocking.
940 dispatchInsts(tid);
941
942 ++iewUnblockCycles;
943
944 if (validInstsFromRename()) {
945 // Add the current inputs to the skid buffer so they can be
946 // reprocessed when this stage unblocks.
947 skidInsert(tid);
948 }
949
950 unblock(tid);
951 }
952 }
953
954 template <class Impl>
955 void
956 DefaultIEW<Impl>::dispatchInsts(ThreadID tid)
957 {
958 // Obtain instructions from skid buffer if unblocking, or queue from rename
959 // otherwise.
960 std::queue<DynInstPtr> &insts_to_dispatch =
961 dispatchStatus[tid] == Unblocking ?
962 skidBuffer[tid] : insts[tid];
963
964 int insts_to_add = insts_to_dispatch.size();
965
966 DynInstPtr inst;
967 bool add_to_iq = false;
968 int dis_num_inst = 0;
969
970 // Loop through the instructions, putting them in the instruction
971 // queue.
972 for ( ; dis_num_inst < insts_to_add &&
973 dis_num_inst < dispatchWidth;
974 ++dis_num_inst)
975 {
976 inst = insts_to_dispatch.front();
977
978 if (dispatchStatus[tid] == Unblocking) {
979 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
980 "buffer\n", tid);
981 }
982
983 // Make sure there's a valid instruction there.
984 assert(inst);
985
986 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %s [sn:%lli] [tid:%i] to "
987 "IQ.\n",
988 tid, inst->pcState(), inst->seqNum, inst->threadNumber);
989
990 // Be sure to mark these instructions as ready so that the
991 // commit stage can go ahead and execute them, and mark
992 // them as issued so the IQ doesn't reprocess them.
993
994 // Check for squashed instructions.
995 if (inst->isSquashed()) {
996 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
997 "not adding to IQ.\n", tid);
998
999 ++iewDispSquashedInsts;
1000
1001 insts_to_dispatch.pop();
1002
1003 //Tell Rename That An Instruction has been processed
1004 if (inst->isLoad()) {
1005 toRename->iewInfo[tid].dispatchedToLQ++;
1006 }
1007 if (inst->isStore()) {
1008 toRename->iewInfo[tid].dispatchedToSQ++;
1009 }
1010
1011 toRename->iewInfo[tid].dispatched++;
1012
1013 continue;
1014 }
1015
1016 // Check for full conditions.
1017 if (instQueue.isFull(tid)) {
1018 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1019
1020 // Call function to start blocking.
1021 block(tid);
1022
1023 // Set unblock to false. Special case where we are using
1024 // skidbuffer (unblocking) instructions but then we still
1025 // get full in the IQ.
1026 toRename->iewUnblock[tid] = false;
1027
1028 ++iewIQFullEvents;
1029 break;
1030 }
1031
1032 // Check LSQ if inst is LD/ST
1033 if ((inst->isLoad() && ldstQueue.lqFull(tid)) ||
1034 (inst->isStore() && ldstQueue.sqFull(tid))) {
1035 DPRINTF(IEW, "[tid:%i]: Issue: %s has become full.\n",tid,
1036 inst->isLoad() ? "LQ" : "SQ");
1037
1038 // Call function to start blocking.
1039 block(tid);
1040
1041 // Set unblock to false. Special case where we are using
1042 // skidbuffer (unblocking) instructions but then we still
1043 // get full in the IQ.
1044 toRename->iewUnblock[tid] = false;
1045
1046 ++iewLSQFullEvents;
1047 break;
1048 }
1049
1050 // Otherwise issue the instruction just fine.
1051 if (inst->isLoad()) {
1052 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1053 "encountered, adding to LSQ.\n", tid);
1054
1055 // Reserve a spot in the load store queue for this
1056 // memory access.
1057 ldstQueue.insertLoad(inst);
1058
1059 ++iewDispLoadInsts;
1060
1061 add_to_iq = true;
1062
1063 toRename->iewInfo[tid].dispatchedToLQ++;
1064 } else if (inst->isStore()) {
1065 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1066 "encountered, adding to LSQ.\n", tid);
1067
1068 ldstQueue.insertStore(inst);
1069
1070 ++iewDispStoreInsts;
1071
1072 if (inst->isStoreConditional()) {
1073 // Store conditionals need to be set as "canCommit()"
1074 // so that commit can process them when they reach the
1075 // head of commit.
1076 // @todo: This is somewhat specific to Alpha.
1077 inst->setCanCommit();
1078 instQueue.insertNonSpec(inst);
1079 add_to_iq = false;
1080
1081 ++iewDispNonSpecInsts;
1082 } else {
1083 add_to_iq = true;
1084 }
1085
1086 toRename->iewInfo[tid].dispatchedToSQ++;
1087 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1088 // Same as non-speculative stores.
1089 inst->setCanCommit();
1090 instQueue.insertBarrier(inst);
1091 add_to_iq = false;
1092 } else if (inst->isNop()) {
1093 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1094 "skipping.\n", tid);
1095
1096 inst->setIssued();
1097 inst->setExecuted();
1098 inst->setCanCommit();
1099
1100 instQueue.recordProducer(inst);
1101
1102 iewExecutedNop[tid]++;
1103
1104 add_to_iq = false;
1105 } else {
1106 assert(!inst->isExecuted());
1107 add_to_iq = true;
1108 }
1109
1110 if (add_to_iq && inst->isNonSpeculative()) {
1111 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1112 "encountered, skipping.\n", tid);
1113
1114 // Same as non-speculative stores.
1115 inst->setCanCommit();
1116
1117 // Specifically insert it as nonspeculative.
1118 instQueue.insertNonSpec(inst);
1119
1120 ++iewDispNonSpecInsts;
1121
1122 add_to_iq = false;
1123 }
1124
1125 // If the instruction queue is not full, then add the
1126 // instruction.
1127 if (add_to_iq) {
1128 instQueue.insert(inst);
1129 }
1130
1131 insts_to_dispatch.pop();
1132
1133 toRename->iewInfo[tid].dispatched++;
1134
1135 ++iewDispatchedInsts;
1136
1137 #if TRACING_ON
1138 inst->dispatchTick = curTick() - inst->fetchTick;
1139 #endif
1140 ppDispatch->notify(inst);
1141 }
1142
1143 if (!insts_to_dispatch.empty()) {
1144 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1145 block(tid);
1146 toRename->iewUnblock[tid] = false;
1147 }
1148
1149 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1150 dispatchStatus[tid] = Running;
1151
1152 updatedQueues = true;
1153 }
1154
1155 dis_num_inst = 0;
1156 }
1157
1158 template <class Impl>
1159 void
1160 DefaultIEW<Impl>::printAvailableInsts()
1161 {
1162 int inst = 0;
1163
1164 std::cout << "Available Instructions: ";
1165
1166 while (fromIssue->insts[inst]) {
1167
1168 if (inst%3==0) std::cout << "\n\t";
1169
1170 std::cout << "PC: " << fromIssue->insts[inst]->pcState()
1171 << " TN: " << fromIssue->insts[inst]->threadNumber
1172 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1173
1174 inst++;
1175
1176 }
1177
1178 std::cout << "\n";
1179 }
1180
1181 template <class Impl>
1182 void
1183 DefaultIEW<Impl>::executeInsts()
1184 {
1185 wbNumInst = 0;
1186 wbCycle = 0;
1187
1188 list<ThreadID>::iterator threads = activeThreads->begin();
1189 list<ThreadID>::iterator end = activeThreads->end();
1190
1191 while (threads != end) {
1192 ThreadID tid = *threads++;
1193 fetchRedirect[tid] = false;
1194 }
1195
1196 // Uncomment this if you want to see all available instructions.
1197 // @todo This doesn't actually work anymore, we should fix it.
1198 // printAvailableInsts();
1199
1200 // Execute/writeback any instructions that are available.
1201 int insts_to_execute = fromIssue->size;
1202 int inst_num = 0;
1203 for (; inst_num < insts_to_execute;
1204 ++inst_num) {
1205
1206 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1207
1208 DynInstPtr inst = instQueue.getInstToExecute();
1209
1210 DPRINTF(IEW, "Execute: Processing PC %s, [tid:%i] [sn:%i].\n",
1211 inst->pcState(), inst->threadNumber,inst->seqNum);
1212
1213 // Notify potential listeners that this instruction has started
1214 // executing
1215 ppExecute->notify(inst);
1216
1217 // Check if the instruction is squashed; if so then skip it
1218 if (inst->isSquashed()) {
1219 DPRINTF(IEW, "Execute: Instruction was squashed. PC: %s, [tid:%i]"
1220 " [sn:%i]\n", inst->pcState(), inst->threadNumber,
1221 inst->seqNum);
1222
1223 // Consider this instruction executed so that commit can go
1224 // ahead and retire the instruction.
1225 inst->setExecuted();
1226
1227 // Not sure if I should set this here or just let commit try to
1228 // commit any squashed instructions. I like the latter a bit more.
1229 inst->setCanCommit();
1230
1231 ++iewExecSquashedInsts;
1232
1233 continue;
1234 }
1235
1236 Fault fault = NoFault;
1237
1238 // Execute instruction.
1239 // Note that if the instruction faults, it will be handled
1240 // at the commit stage.
1241 if (inst->isMemRef()) {
1242 DPRINTF(IEW, "Execute: Calculating address for memory "
1243 "reference.\n");
1244
1245 // Tell the LDSTQ to execute this instruction (if it is a load).
1246 if (inst->isLoad()) {
1247 // Loads will mark themselves as executed, and their writeback
1248 // event adds the instruction to the queue to commit
1249 fault = ldstQueue.executeLoad(inst);
1250
1251 if (inst->isTranslationDelayed() &&
1252 fault == NoFault) {
1253 // A hw page table walk is currently going on; the
1254 // instruction must be deferred.
1255 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1256 "load.\n");
1257 instQueue.deferMemInst(inst);
1258 continue;
1259 }
1260
1261 if (inst->isDataPrefetch() || inst->isInstPrefetch()) {
1262 inst->fault = NoFault;
1263 }
1264 } else if (inst->isStore()) {
1265 fault = ldstQueue.executeStore(inst);
1266
1267 if (inst->isTranslationDelayed() &&
1268 fault == NoFault) {
1269 // A hw page table walk is currently going on; the
1270 // instruction must be deferred.
1271 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1272 "store.\n");
1273 instQueue.deferMemInst(inst);
1274 continue;
1275 }
1276
1277 // If the store had a fault then it may not have a mem req
1278 if (fault != NoFault || !inst->readPredicate() ||
1279 !inst->isStoreConditional()) {
1280 // If the instruction faulted, then we need to send it along
1281 // to commit without the instruction completing.
1282 // Send this instruction to commit, also make sure iew stage
1283 // realizes there is activity.
1284 inst->setExecuted();
1285 instToCommit(inst);
1286 activityThisCycle();
1287 }
1288
1289 // Store conditionals will mark themselves as
1290 // executed, and their writeback event will add the
1291 // instruction to the queue to commit.
1292 } else {
1293 panic("Unexpected memory type!\n");
1294 }
1295
1296 } else {
1297 // If the instruction has already faulted, then skip executing it.
1298 // Such case can happen when it faulted during ITLB translation.
1299 // If we execute the instruction (even if it's a nop) the fault
1300 // will be replaced and we will lose it.
1301 if (inst->getFault() == NoFault) {
1302 inst->execute();
1303 if (!inst->readPredicate())
1304 inst->forwardOldRegs();
1305 }
1306
1307 inst->setExecuted();
1308
1309 instToCommit(inst);
1310 }
1311
1312 updateExeInstStats(inst);
1313
1314 // Check if branch prediction was correct, if not then we need
1315 // to tell commit to squash in flight instructions. Only
1316 // handle this if there hasn't already been something that
1317 // redirects fetch in this group of instructions.
1318
1319 // This probably needs to prioritize the redirects if a different
1320 // scheduler is used. Currently the scheduler schedules the oldest
1321 // instruction first, so the branch resolution order will be correct.
1322 ThreadID tid = inst->threadNumber;
1323
1324 if (!fetchRedirect[tid] ||
1325 !toCommit->squash[tid] ||
1326 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1327
1328 // Prevent testing for misprediction on load instructions,
1329 // that have not been executed.
1330 bool loadNotExecuted = !inst->isExecuted() && inst->isLoad();
1331
1332 if (inst->mispredicted() && !loadNotExecuted) {
1333 fetchRedirect[tid] = true;
1334
1335 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1336 DPRINTF(IEW, "Predicted target was PC: %s.\n",
1337 inst->readPredTarg());
1338 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %s.\n",
1339 inst->pcState());
1340 // If incorrect, then signal the ROB that it must be squashed.
1341 squashDueToBranch(inst, tid);
1342
1343 ppMispredict->notify(inst);
1344
1345 if (inst->readPredTaken()) {
1346 predictedTakenIncorrect++;
1347 } else {
1348 predictedNotTakenIncorrect++;
1349 }
1350 } else if (ldstQueue.violation(tid)) {
1351 assert(inst->isMemRef());
1352 // If there was an ordering violation, then get the
1353 // DynInst that caused the violation. Note that this
1354 // clears the violation signal.
1355 DynInstPtr violator;
1356 violator = ldstQueue.getMemDepViolator(tid);
1357
1358 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: %s "
1359 "[sn:%lli], inst PC: %s [sn:%lli]. Addr is: %#x.\n",
1360 violator->pcState(), violator->seqNum,
1361 inst->pcState(), inst->seqNum, inst->physEffAddr);
1362
1363 fetchRedirect[tid] = true;
1364
1365 // Tell the instruction queue that a violation has occured.
1366 instQueue.violation(inst, violator);
1367
1368 // Squash.
1369 squashDueToMemOrder(violator, tid);
1370
1371 ++memOrderViolationEvents;
1372 }
1373 } else {
1374 // Reset any state associated with redirects that will not
1375 // be used.
1376 if (ldstQueue.violation(tid)) {
1377 assert(inst->isMemRef());
1378
1379 DynInstPtr violator = ldstQueue.getMemDepViolator(tid);
1380
1381 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1382 "%s, inst PC: %s. Addr is: %#x.\n",
1383 violator->pcState(), inst->pcState(),
1384 inst->physEffAddr);
1385 DPRINTF(IEW, "Violation will not be handled because "
1386 "already squashing\n");
1387
1388 ++memOrderViolationEvents;
1389 }
1390 }
1391 }
1392
1393 // Update and record activity if we processed any instructions.
1394 if (inst_num) {
1395 if (exeStatus == Idle) {
1396 exeStatus = Running;
1397 }
1398
1399 updatedQueues = true;
1400
1401 cpu->activityThisCycle();
1402 }
1403
1404 // Need to reset this in case a writeback event needs to write into the
1405 // iew queue. That way the writeback event will write into the correct
1406 // spot in the queue.
1407 wbNumInst = 0;
1408
1409 }
1410
1411 template <class Impl>
1412 void
1413 DefaultIEW<Impl>::writebackInsts()
1414 {
1415 // Loop through the head of the time buffer and wake any
1416 // dependents. These instructions are about to write back. Also
1417 // mark scoreboard that this instruction is finally complete.
1418 // Either have IEW have direct access to scoreboard, or have this
1419 // as part of backwards communication.
1420 for (int inst_num = 0; inst_num < wbWidth &&
1421 toCommit->insts[inst_num]; inst_num++) {
1422 DynInstPtr inst = toCommit->insts[inst_num];
1423 ThreadID tid = inst->threadNumber;
1424
1425 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %s.\n",
1426 inst->seqNum, inst->pcState());
1427
1428 iewInstsToCommit[tid]++;
1429 // Notify potential listeners that execution is complete for this
1430 // instruction.
1431 ppToCommit->notify(inst);
1432
1433 // Some instructions will be sent to commit without having
1434 // executed because they need commit to handle them.
1435 // E.g. Strictly ordered loads have not actually executed when they
1436 // are first sent to commit. Instead commit must tell the LSQ
1437 // when it's ready to execute the strictly ordered load.
1438 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1439 int dependents = instQueue.wakeDependents(inst);
1440
1441 for (int i = 0; i < inst->numDestRegs(); i++) {
1442 //mark as Ready
1443 DPRINTF(IEW,"Setting Destination Register %i (%s)\n",
1444 inst->renamedDestRegIdx(i)->index(),
1445 inst->renamedDestRegIdx(i)->className());
1446 scoreboard->setReg(inst->renamedDestRegIdx(i));
1447 }
1448
1449 if (dependents) {
1450 producerInst[tid]++;
1451 consumerInst[tid]+= dependents;
1452 }
1453 writebackCount[tid]++;
1454 }
1455 }
1456 }
1457
1458 template<class Impl>
1459 void
1460 DefaultIEW<Impl>::tick()
1461 {
1462 wbNumInst = 0;
1463 wbCycle = 0;
1464
1465 wroteToTimeBuffer = false;
1466 updatedQueues = false;
1467
1468 ldstQueue.tick();
1469
1470 sortInsts();
1471
1472 // Free function units marked as being freed this cycle.
1473 fuPool->processFreeUnits();
1474
1475 list<ThreadID>::iterator threads = activeThreads->begin();
1476 list<ThreadID>::iterator end = activeThreads->end();
1477
1478 // Check stall and squash signals, dispatch any instructions.
1479 while (threads != end) {
1480 ThreadID tid = *threads++;
1481
1482 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1483
1484 checkSignalsAndUpdate(tid);
1485 dispatch(tid);
1486 }
1487
1488 if (exeStatus != Squashing) {
1489 executeInsts();
1490
1491 writebackInsts();
1492
1493 // Have the instruction queue try to schedule any ready instructions.
1494 // (In actuality, this scheduling is for instructions that will
1495 // be executed next cycle.)
1496 instQueue.scheduleReadyInsts();
1497
1498 // Also should advance its own time buffers if the stage ran.
1499 // Not the best place for it, but this works (hopefully).
1500 issueToExecQueue.advance();
1501 }
1502
1503 bool broadcast_free_entries = false;
1504
1505 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1506 exeStatus = Idle;
1507 updateLSQNextCycle = false;
1508
1509 broadcast_free_entries = true;
1510 }
1511
1512 // Writeback any stores using any leftover bandwidth.
1513 ldstQueue.writebackStores();
1514
1515 // Check the committed load/store signals to see if there's a load
1516 // or store to commit. Also check if it's being told to execute a
1517 // nonspeculative instruction.
1518 // This is pretty inefficient...
1519
1520 threads = activeThreads->begin();
1521 while (threads != end) {
1522 ThreadID tid = (*threads++);
1523
1524 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1525
1526 // Update structures based on instructions committed.
1527 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1528 !fromCommit->commitInfo[tid].squash &&
1529 !fromCommit->commitInfo[tid].robSquashing) {
1530
1531 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1532
1533 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1534
1535 updateLSQNextCycle = true;
1536 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1537 }
1538
1539 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1540
1541 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1542 if (fromCommit->commitInfo[tid].strictlyOrdered) {
1543 instQueue.replayMemInst(
1544 fromCommit->commitInfo[tid].strictlyOrderedLoad);
1545 fromCommit->commitInfo[tid].strictlyOrderedLoad->setAtCommit();
1546 } else {
1547 instQueue.scheduleNonSpec(
1548 fromCommit->commitInfo[tid].nonSpecSeqNum);
1549 }
1550 }
1551
1552 if (broadcast_free_entries) {
1553 toFetch->iewInfo[tid].iqCount =
1554 instQueue.getCount(tid);
1555 toFetch->iewInfo[tid].ldstqCount =
1556 ldstQueue.getCount(tid);
1557
1558 toRename->iewInfo[tid].usedIQ = true;
1559 toRename->iewInfo[tid].freeIQEntries =
1560 instQueue.numFreeEntries(tid);
1561 toRename->iewInfo[tid].usedLSQ = true;
1562
1563 toRename->iewInfo[tid].freeLQEntries =
1564 ldstQueue.numFreeLoadEntries(tid);
1565 toRename->iewInfo[tid].freeSQEntries =
1566 ldstQueue.numFreeStoreEntries(tid);
1567
1568 wroteToTimeBuffer = true;
1569 }
1570
1571 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1572 tid, toRename->iewInfo[tid].dispatched);
1573 }
1574
1575 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1576 "LQ has %i free entries. SQ has %i free entries.\n",
1577 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1578 ldstQueue.numFreeLoadEntries(), ldstQueue.numFreeStoreEntries());
1579
1580 updateStatus();
1581
1582 if (wroteToTimeBuffer) {
1583 DPRINTF(Activity, "Activity this cycle.\n");
1584 cpu->activityThisCycle();
1585 }
1586 }
1587
1588 template <class Impl>
1589 void
1590 DefaultIEW<Impl>::updateExeInstStats(const DynInstPtr& inst)
1591 {
1592 ThreadID tid = inst->threadNumber;
1593
1594 iewExecutedInsts++;
1595
1596 #if TRACING_ON
1597 if (DTRACE(O3PipeView)) {
1598 inst->completeTick = curTick() - inst->fetchTick;
1599 }
1600 #endif
1601
1602 //
1603 // Control operations
1604 //
1605 if (inst->isControl())
1606 iewExecutedBranches[tid]++;
1607
1608 //
1609 // Memory operations
1610 //
1611 if (inst->isMemRef()) {
1612 iewExecutedRefs[tid]++;
1613
1614 if (inst->isLoad()) {
1615 iewExecLoadInsts[tid]++;
1616 }
1617 }
1618 }
1619
1620 template <class Impl>
1621 void
1622 DefaultIEW<Impl>::checkMisprediction(const DynInstPtr& inst)
1623 {
1624 ThreadID tid = inst->threadNumber;
1625
1626 if (!fetchRedirect[tid] ||
1627 !toCommit->squash[tid] ||
1628 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1629
1630 if (inst->mispredicted()) {
1631 fetchRedirect[tid] = true;
1632
1633 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1634 DPRINTF(IEW, "Predicted target was PC:%#x, NPC:%#x.\n",
1635 inst->predInstAddr(), inst->predNextInstAddr());
1636 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x,"
1637 " NPC: %#x.\n", inst->nextInstAddr(),
1638 inst->nextInstAddr());
1639 // If incorrect, then signal the ROB that it must be squashed.
1640 squashDueToBranch(inst, tid);
1641
1642 if (inst->readPredTaken()) {
1643 predictedTakenIncorrect++;
1644 } else {
1645 predictedNotTakenIncorrect++;
1646 }
1647 }
1648 }
1649 }
1650
1651 #endif//__CPU_O3_IEW_IMPL_IMPL_HH__