iew_impl.hh (9444:ab47fe7f03f0) iew_impl.hh (9527:68154bc0e0ea)
1/*
2 * Copyright (c) 2010-2012 ARM Limited
3 * All rights reserved.
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2004-2006 The Regents of The University of Michigan
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Kevin Lim
41 */
42
43// @todo: Fix the instantaneous communication among all the stages within
44// iew. There's a clear delay between issue and execute, yet backwards
45// communication happens simultaneously.
46
47#include <queue>
48
49#include "arch/utility.hh"
50#include "config/the_isa.hh"
51#include "cpu/checker/cpu.hh"
52#include "cpu/o3/fu_pool.hh"
53#include "cpu/o3/iew.hh"
54#include "cpu/timebuf.hh"
55#include "debug/Activity.hh"
56#include "debug/Decode.hh"
57#include "debug/Drain.hh"
58#include "debug/IEW.hh"
1/*
2 * Copyright (c) 2010-2012 ARM Limited
3 * All rights reserved.
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2004-2006 The Regents of The University of Michigan
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Kevin Lim
41 */
42
43// @todo: Fix the instantaneous communication among all the stages within
44// iew. There's a clear delay between issue and execute, yet backwards
45// communication happens simultaneously.
46
47#include <queue>
48
49#include "arch/utility.hh"
50#include "config/the_isa.hh"
51#include "cpu/checker/cpu.hh"
52#include "cpu/o3/fu_pool.hh"
53#include "cpu/o3/iew.hh"
54#include "cpu/timebuf.hh"
55#include "debug/Activity.hh"
56#include "debug/Decode.hh"
57#include "debug/Drain.hh"
58#include "debug/IEW.hh"
59#include "debug/O3PipeView.hh"
59#include "params/DerivO3CPU.hh"
60
61using namespace std;
62
63template<class Impl>
64DefaultIEW<Impl>::DefaultIEW(O3CPU *_cpu, DerivO3CPUParams *params)
65 : issueToExecQueue(params->backComSize, params->forwardComSize),
66 cpu(_cpu),
67 instQueue(_cpu, this, params),
68 ldstQueue(_cpu, this, params),
69 fuPool(params->fuPool),
70 commitToIEWDelay(params->commitToIEWDelay),
71 renameToIEWDelay(params->renameToIEWDelay),
72 issueToExecuteDelay(params->issueToExecuteDelay),
73 dispatchWidth(params->dispatchWidth),
74 issueWidth(params->issueWidth),
75 wbOutstanding(0),
76 wbWidth(params->wbWidth),
77 numThreads(params->numThreads)
78{
79 _status = Active;
80 exeStatus = Running;
81 wbStatus = Idle;
82
83 // Setup wire to read instructions coming from issue.
84 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
85
86 // Instruction queue needs the queue between issue and execute.
87 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
88
89 for (ThreadID tid = 0; tid < numThreads; tid++) {
90 dispatchStatus[tid] = Running;
91 stalls[tid].commit = false;
92 fetchRedirect[tid] = false;
93 }
94
95 wbMax = wbWidth * params->wbDepth;
96
97 updateLSQNextCycle = false;
98
99 ableToIssue = true;
100
101 skidBufferMax = (3 * (renameToIEWDelay * params->renameWidth)) + issueWidth;
102}
103
104template <class Impl>
105std::string
106DefaultIEW<Impl>::name() const
107{
108 return cpu->name() + ".iew";
109}
110
111template <class Impl>
112void
113DefaultIEW<Impl>::regStats()
114{
115 using namespace Stats;
116
117 instQueue.regStats();
118 ldstQueue.regStats();
119
120 iewIdleCycles
121 .name(name() + ".iewIdleCycles")
122 .desc("Number of cycles IEW is idle");
123
124 iewSquashCycles
125 .name(name() + ".iewSquashCycles")
126 .desc("Number of cycles IEW is squashing");
127
128 iewBlockCycles
129 .name(name() + ".iewBlockCycles")
130 .desc("Number of cycles IEW is blocking");
131
132 iewUnblockCycles
133 .name(name() + ".iewUnblockCycles")
134 .desc("Number of cycles IEW is unblocking");
135
136 iewDispatchedInsts
137 .name(name() + ".iewDispatchedInsts")
138 .desc("Number of instructions dispatched to IQ");
139
140 iewDispSquashedInsts
141 .name(name() + ".iewDispSquashedInsts")
142 .desc("Number of squashed instructions skipped by dispatch");
143
144 iewDispLoadInsts
145 .name(name() + ".iewDispLoadInsts")
146 .desc("Number of dispatched load instructions");
147
148 iewDispStoreInsts
149 .name(name() + ".iewDispStoreInsts")
150 .desc("Number of dispatched store instructions");
151
152 iewDispNonSpecInsts
153 .name(name() + ".iewDispNonSpecInsts")
154 .desc("Number of dispatched non-speculative instructions");
155
156 iewIQFullEvents
157 .name(name() + ".iewIQFullEvents")
158 .desc("Number of times the IQ has become full, causing a stall");
159
160 iewLSQFullEvents
161 .name(name() + ".iewLSQFullEvents")
162 .desc("Number of times the LSQ has become full, causing a stall");
163
164 memOrderViolationEvents
165 .name(name() + ".memOrderViolationEvents")
166 .desc("Number of memory order violations");
167
168 predictedTakenIncorrect
169 .name(name() + ".predictedTakenIncorrect")
170 .desc("Number of branches that were predicted taken incorrectly");
171
172 predictedNotTakenIncorrect
173 .name(name() + ".predictedNotTakenIncorrect")
174 .desc("Number of branches that were predicted not taken incorrectly");
175
176 branchMispredicts
177 .name(name() + ".branchMispredicts")
178 .desc("Number of branch mispredicts detected at execute");
179
180 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
181
182 iewExecutedInsts
183 .name(name() + ".iewExecutedInsts")
184 .desc("Number of executed instructions");
185
186 iewExecLoadInsts
187 .init(cpu->numThreads)
188 .name(name() + ".iewExecLoadInsts")
189 .desc("Number of load instructions executed")
190 .flags(total);
191
192 iewExecSquashedInsts
193 .name(name() + ".iewExecSquashedInsts")
194 .desc("Number of squashed instructions skipped in execute");
195
196 iewExecutedSwp
197 .init(cpu->numThreads)
198 .name(name() + ".exec_swp")
199 .desc("number of swp insts executed")
200 .flags(total);
201
202 iewExecutedNop
203 .init(cpu->numThreads)
204 .name(name() + ".exec_nop")
205 .desc("number of nop insts executed")
206 .flags(total);
207
208 iewExecutedRefs
209 .init(cpu->numThreads)
210 .name(name() + ".exec_refs")
211 .desc("number of memory reference insts executed")
212 .flags(total);
213
214 iewExecutedBranches
215 .init(cpu->numThreads)
216 .name(name() + ".exec_branches")
217 .desc("Number of branches executed")
218 .flags(total);
219
220 iewExecStoreInsts
221 .name(name() + ".exec_stores")
222 .desc("Number of stores executed")
223 .flags(total);
224 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
225
226 iewExecRate
227 .name(name() + ".exec_rate")
228 .desc("Inst execution rate")
229 .flags(total);
230
231 iewExecRate = iewExecutedInsts / cpu->numCycles;
232
233 iewInstsToCommit
234 .init(cpu->numThreads)
235 .name(name() + ".wb_sent")
236 .desc("cumulative count of insts sent to commit")
237 .flags(total);
238
239 writebackCount
240 .init(cpu->numThreads)
241 .name(name() + ".wb_count")
242 .desc("cumulative count of insts written-back")
243 .flags(total);
244
245 producerInst
246 .init(cpu->numThreads)
247 .name(name() + ".wb_producers")
248 .desc("num instructions producing a value")
249 .flags(total);
250
251 consumerInst
252 .init(cpu->numThreads)
253 .name(name() + ".wb_consumers")
254 .desc("num instructions consuming a value")
255 .flags(total);
256
257 wbPenalized
258 .init(cpu->numThreads)
259 .name(name() + ".wb_penalized")
260 .desc("number of instrctions required to write to 'other' IQ")
261 .flags(total);
262
263 wbPenalizedRate
264 .name(name() + ".wb_penalized_rate")
265 .desc ("fraction of instructions written-back that wrote to 'other' IQ")
266 .flags(total);
267
268 wbPenalizedRate = wbPenalized / writebackCount;
269
270 wbFanout
271 .name(name() + ".wb_fanout")
272 .desc("average fanout of values written-back")
273 .flags(total);
274
275 wbFanout = producerInst / consumerInst;
276
277 wbRate
278 .name(name() + ".wb_rate")
279 .desc("insts written-back per cycle")
280 .flags(total);
281 wbRate = writebackCount / cpu->numCycles;
282}
283
284template<class Impl>
285void
286DefaultIEW<Impl>::startupStage()
287{
288 for (ThreadID tid = 0; tid < numThreads; tid++) {
289 toRename->iewInfo[tid].usedIQ = true;
290 toRename->iewInfo[tid].freeIQEntries =
291 instQueue.numFreeEntries(tid);
292
293 toRename->iewInfo[tid].usedLSQ = true;
294 toRename->iewInfo[tid].freeLSQEntries =
295 ldstQueue.numFreeEntries(tid);
296 }
297
298 // Initialize the checker's dcache port here
299 if (cpu->checker) {
300 cpu->checker->setDcachePort(&cpu->getDataPort());
301 }
302
303 cpu->activateStage(O3CPU::IEWIdx);
304}
305
306template<class Impl>
307void
308DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
309{
310 timeBuffer = tb_ptr;
311
312 // Setup wire to read information from time buffer, from commit.
313 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
314
315 // Setup wire to write information back to previous stages.
316 toRename = timeBuffer->getWire(0);
317
318 toFetch = timeBuffer->getWire(0);
319
320 // Instruction queue also needs main time buffer.
321 instQueue.setTimeBuffer(tb_ptr);
322}
323
324template<class Impl>
325void
326DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
327{
328 renameQueue = rq_ptr;
329
330 // Setup wire to read information from rename queue.
331 fromRename = renameQueue->getWire(-renameToIEWDelay);
332}
333
334template<class Impl>
335void
336DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
337{
338 iewQueue = iq_ptr;
339
340 // Setup wire to write instructions to commit.
341 toCommit = iewQueue->getWire(0);
342}
343
344template<class Impl>
345void
346DefaultIEW<Impl>::setActiveThreads(list<ThreadID> *at_ptr)
347{
348 activeThreads = at_ptr;
349
350 ldstQueue.setActiveThreads(at_ptr);
351 instQueue.setActiveThreads(at_ptr);
352}
353
354template<class Impl>
355void
356DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
357{
358 scoreboard = sb_ptr;
359}
360
361template <class Impl>
362bool
363DefaultIEW<Impl>::isDrained() const
364{
365 bool drained(ldstQueue.isDrained());
366
367 for (ThreadID tid = 0; tid < numThreads; tid++) {
368 if (!insts[tid].empty()) {
369 DPRINTF(Drain, "%i: Insts not empty.\n", tid);
370 drained = false;
371 }
372 if (!skidBuffer[tid].empty()) {
373 DPRINTF(Drain, "%i: Skid buffer not empty.\n", tid);
374 drained = false;
375 }
376 }
377
378 return drained;
379}
380
381template <class Impl>
382void
383DefaultIEW<Impl>::drainSanityCheck() const
384{
385 assert(isDrained());
386
387 instQueue.drainSanityCheck();
388 ldstQueue.drainSanityCheck();
389 fuPool->drainSanityCheck();
390}
391
392template <class Impl>
393void
394DefaultIEW<Impl>::takeOverFrom()
395{
396 // Reset all state.
397 _status = Active;
398 exeStatus = Running;
399 wbStatus = Idle;
400
401 instQueue.takeOverFrom();
402 ldstQueue.takeOverFrom();
403 fuPool->takeOverFrom();
404
405 startupStage();
406 cpu->activityThisCycle();
407
408 for (ThreadID tid = 0; tid < numThreads; tid++) {
409 dispatchStatus[tid] = Running;
410 stalls[tid].commit = false;
411 fetchRedirect[tid] = false;
412 }
413
414 updateLSQNextCycle = false;
415
416 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
417 issueToExecQueue.advance();
418 }
419}
420
421template<class Impl>
422void
423DefaultIEW<Impl>::squash(ThreadID tid)
424{
425 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n", tid);
426
427 // Tell the IQ to start squashing.
428 instQueue.squash(tid);
429
430 // Tell the LDSTQ to start squashing.
431 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
432 updatedQueues = true;
433
434 // Clear the skid buffer in case it has any data in it.
435 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
436 tid, fromCommit->commitInfo[tid].doneSeqNum);
437
438 while (!skidBuffer[tid].empty()) {
439 if (skidBuffer[tid].front()->isLoad() ||
440 skidBuffer[tid].front()->isStore() ) {
441 toRename->iewInfo[tid].dispatchedToLSQ++;
442 }
443
444 toRename->iewInfo[tid].dispatched++;
445
446 skidBuffer[tid].pop();
447 }
448
449 emptyRenameInsts(tid);
450}
451
452template<class Impl>
453void
454DefaultIEW<Impl>::squashDueToBranch(DynInstPtr &inst, ThreadID tid)
455{
456 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %s "
457 "[sn:%i].\n", tid, inst->pcState(), inst->seqNum);
458
459 if (toCommit->squash[tid] == false ||
460 inst->seqNum < toCommit->squashedSeqNum[tid]) {
461 toCommit->squash[tid] = true;
462 toCommit->squashedSeqNum[tid] = inst->seqNum;
463 toCommit->branchTaken[tid] = inst->pcState().branching();
464
465 TheISA::PCState pc = inst->pcState();
466 TheISA::advancePC(pc, inst->staticInst);
467
468 toCommit->pc[tid] = pc;
469 toCommit->mispredictInst[tid] = inst;
470 toCommit->includeSquashInst[tid] = false;
471
472 wroteToTimeBuffer = true;
473 }
474
475}
476
477template<class Impl>
478void
479DefaultIEW<Impl>::squashDueToMemOrder(DynInstPtr &inst, ThreadID tid)
480{
481 DPRINTF(IEW, "[tid:%i]: Memory violation, squashing violator and younger "
482 "insts, PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
483 // Need to include inst->seqNum in the following comparison to cover the
484 // corner case when a branch misprediction and a memory violation for the
485 // same instruction (e.g. load PC) are detected in the same cycle. In this
486 // case the memory violator should take precedence over the branch
487 // misprediction because it requires the violator itself to be included in
488 // the squash.
489 if (toCommit->squash[tid] == false ||
490 inst->seqNum <= toCommit->squashedSeqNum[tid]) {
491 toCommit->squash[tid] = true;
492
493 toCommit->squashedSeqNum[tid] = inst->seqNum;
494 toCommit->pc[tid] = inst->pcState();
495 toCommit->mispredictInst[tid] = NULL;
496
497 // Must include the memory violator in the squash.
498 toCommit->includeSquashInst[tid] = true;
499
500 wroteToTimeBuffer = true;
501 }
502}
503
504template<class Impl>
505void
506DefaultIEW<Impl>::squashDueToMemBlocked(DynInstPtr &inst, ThreadID tid)
507{
508 DPRINTF(IEW, "[tid:%i]: Memory blocked, squashing load and younger insts, "
509 "PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
510 if (toCommit->squash[tid] == false ||
511 inst->seqNum < toCommit->squashedSeqNum[tid]) {
512 toCommit->squash[tid] = true;
513
514 toCommit->squashedSeqNum[tid] = inst->seqNum;
515 toCommit->pc[tid] = inst->pcState();
516 toCommit->mispredictInst[tid] = NULL;
517
518 // Must include the broadcasted SN in the squash.
519 toCommit->includeSquashInst[tid] = true;
520
521 ldstQueue.setLoadBlockedHandled(tid);
522
523 wroteToTimeBuffer = true;
524 }
525}
526
527template<class Impl>
528void
529DefaultIEW<Impl>::block(ThreadID tid)
530{
531 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
532
533 if (dispatchStatus[tid] != Blocked &&
534 dispatchStatus[tid] != Unblocking) {
535 toRename->iewBlock[tid] = true;
536 wroteToTimeBuffer = true;
537 }
538
539 // Add the current inputs to the skid buffer so they can be
540 // reprocessed when this stage unblocks.
541 skidInsert(tid);
542
543 dispatchStatus[tid] = Blocked;
544}
545
546template<class Impl>
547void
548DefaultIEW<Impl>::unblock(ThreadID tid)
549{
550 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
551 "buffer %u.\n",tid, tid);
552
553 // If the skid bufffer is empty, signal back to previous stages to unblock.
554 // Also switch status to running.
555 if (skidBuffer[tid].empty()) {
556 toRename->iewUnblock[tid] = true;
557 wroteToTimeBuffer = true;
558 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
559 dispatchStatus[tid] = Running;
560 }
561}
562
563template<class Impl>
564void
565DefaultIEW<Impl>::wakeDependents(DynInstPtr &inst)
566{
567 instQueue.wakeDependents(inst);
568}
569
570template<class Impl>
571void
572DefaultIEW<Impl>::rescheduleMemInst(DynInstPtr &inst)
573{
574 instQueue.rescheduleMemInst(inst);
575}
576
577template<class Impl>
578void
579DefaultIEW<Impl>::replayMemInst(DynInstPtr &inst)
580{
581 instQueue.replayMemInst(inst);
582}
583
584template<class Impl>
585void
586DefaultIEW<Impl>::instToCommit(DynInstPtr &inst)
587{
588 // This function should not be called after writebackInsts in a
589 // single cycle. That will cause problems with an instruction
590 // being added to the queue to commit without being processed by
591 // writebackInsts prior to being sent to commit.
592
593 // First check the time slot that this instruction will write
594 // to. If there are free write ports at the time, then go ahead
595 // and write the instruction to that time. If there are not,
596 // keep looking back to see where's the first time there's a
597 // free slot.
598 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
599 ++wbNumInst;
600 if (wbNumInst == wbWidth) {
601 ++wbCycle;
602 wbNumInst = 0;
603 }
604
605 assert((wbCycle * wbWidth + wbNumInst) <= wbMax);
606 }
607
608 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
609 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
610 // Add finished instruction to queue to commit.
611 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
612 (*iewQueue)[wbCycle].size++;
613}
614
615template <class Impl>
616unsigned
617DefaultIEW<Impl>::validInstsFromRename()
618{
619 unsigned inst_count = 0;
620
621 for (int i=0; i<fromRename->size; i++) {
622 if (!fromRename->insts[i]->isSquashed())
623 inst_count++;
624 }
625
626 return inst_count;
627}
628
629template<class Impl>
630void
631DefaultIEW<Impl>::skidInsert(ThreadID tid)
632{
633 DynInstPtr inst = NULL;
634
635 while (!insts[tid].empty()) {
636 inst = insts[tid].front();
637
638 insts[tid].pop();
639
640 DPRINTF(Decode,"[tid:%i]: Inserting [sn:%lli] PC:%s into "
641 "dispatch skidBuffer %i\n",tid, inst->seqNum,
642 inst->pcState(),tid);
643
644 skidBuffer[tid].push(inst);
645 }
646
647 assert(skidBuffer[tid].size() <= skidBufferMax &&
648 "Skidbuffer Exceeded Max Size");
649}
650
651template<class Impl>
652int
653DefaultIEW<Impl>::skidCount()
654{
655 int max=0;
656
657 list<ThreadID>::iterator threads = activeThreads->begin();
658 list<ThreadID>::iterator end = activeThreads->end();
659
660 while (threads != end) {
661 ThreadID tid = *threads++;
662 unsigned thread_count = skidBuffer[tid].size();
663 if (max < thread_count)
664 max = thread_count;
665 }
666
667 return max;
668}
669
670template<class Impl>
671bool
672DefaultIEW<Impl>::skidsEmpty()
673{
674 list<ThreadID>::iterator threads = activeThreads->begin();
675 list<ThreadID>::iterator end = activeThreads->end();
676
677 while (threads != end) {
678 ThreadID tid = *threads++;
679
680 if (!skidBuffer[tid].empty())
681 return false;
682 }
683
684 return true;
685}
686
687template <class Impl>
688void
689DefaultIEW<Impl>::updateStatus()
690{
691 bool any_unblocking = false;
692
693 list<ThreadID>::iterator threads = activeThreads->begin();
694 list<ThreadID>::iterator end = activeThreads->end();
695
696 while (threads != end) {
697 ThreadID tid = *threads++;
698
699 if (dispatchStatus[tid] == Unblocking) {
700 any_unblocking = true;
701 break;
702 }
703 }
704
705 // If there are no ready instructions waiting to be scheduled by the IQ,
706 // and there's no stores waiting to write back, and dispatch is not
707 // unblocking, then there is no internal activity for the IEW stage.
708 instQueue.intInstQueueReads++;
709 if (_status == Active && !instQueue.hasReadyInsts() &&
710 !ldstQueue.willWB() && !any_unblocking) {
711 DPRINTF(IEW, "IEW switching to idle\n");
712
713 deactivateStage();
714
715 _status = Inactive;
716 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
717 ldstQueue.willWB() ||
718 any_unblocking)) {
719 // Otherwise there is internal activity. Set to active.
720 DPRINTF(IEW, "IEW switching to active\n");
721
722 activateStage();
723
724 _status = Active;
725 }
726}
727
728template <class Impl>
729void
730DefaultIEW<Impl>::resetEntries()
731{
732 instQueue.resetEntries();
733 ldstQueue.resetEntries();
734}
735
736template <class Impl>
737void
738DefaultIEW<Impl>::readStallSignals(ThreadID tid)
739{
740 if (fromCommit->commitBlock[tid]) {
741 stalls[tid].commit = true;
742 }
743
744 if (fromCommit->commitUnblock[tid]) {
745 assert(stalls[tid].commit);
746 stalls[tid].commit = false;
747 }
748}
749
750template <class Impl>
751bool
752DefaultIEW<Impl>::checkStall(ThreadID tid)
753{
754 bool ret_val(false);
755
756 if (stalls[tid].commit) {
757 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
758 ret_val = true;
759 } else if (instQueue.isFull(tid)) {
760 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
761 ret_val = true;
762 } else if (ldstQueue.isFull(tid)) {
763 DPRINTF(IEW,"[tid:%i]: Stall: LSQ is full\n",tid);
764
765 if (ldstQueue.numLoads(tid) > 0 ) {
766
767 DPRINTF(IEW,"[tid:%i]: LSQ oldest load: [sn:%i] \n",
768 tid,ldstQueue.getLoadHeadSeqNum(tid));
769 }
770
771 if (ldstQueue.numStores(tid) > 0) {
772
773 DPRINTF(IEW,"[tid:%i]: LSQ oldest store: [sn:%i] \n",
774 tid,ldstQueue.getStoreHeadSeqNum(tid));
775 }
776
777 ret_val = true;
778 } else if (ldstQueue.isStalled(tid)) {
779 DPRINTF(IEW,"[tid:%i]: Stall: LSQ stall detected.\n",tid);
780 ret_val = true;
781 }
782
783 return ret_val;
784}
785
786template <class Impl>
787void
788DefaultIEW<Impl>::checkSignalsAndUpdate(ThreadID tid)
789{
790 // Check if there's a squash signal, squash if there is
791 // Check stall signals, block if there is.
792 // If status was Blocked
793 // if so then go to unblocking
794 // If status was Squashing
795 // check if squashing is not high. Switch to running this cycle.
796
797 readStallSignals(tid);
798
799 if (fromCommit->commitInfo[tid].squash) {
800 squash(tid);
801
802 if (dispatchStatus[tid] == Blocked ||
803 dispatchStatus[tid] == Unblocking) {
804 toRename->iewUnblock[tid] = true;
805 wroteToTimeBuffer = true;
806 }
807
808 dispatchStatus[tid] = Squashing;
809 fetchRedirect[tid] = false;
810 return;
811 }
812
813 if (fromCommit->commitInfo[tid].robSquashing) {
814 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
815
816 dispatchStatus[tid] = Squashing;
817 emptyRenameInsts(tid);
818 wroteToTimeBuffer = true;
819 return;
820 }
821
822 if (checkStall(tid)) {
823 block(tid);
824 dispatchStatus[tid] = Blocked;
825 return;
826 }
827
828 if (dispatchStatus[tid] == Blocked) {
829 // Status from previous cycle was blocked, but there are no more stall
830 // conditions. Switch over to unblocking.
831 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
832 tid);
833
834 dispatchStatus[tid] = Unblocking;
835
836 unblock(tid);
837
838 return;
839 }
840
841 if (dispatchStatus[tid] == Squashing) {
842 // Switch status to running if rename isn't being told to block or
843 // squash this cycle.
844 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
845 tid);
846
847 dispatchStatus[tid] = Running;
848
849 return;
850 }
851}
852
853template <class Impl>
854void
855DefaultIEW<Impl>::sortInsts()
856{
857 int insts_from_rename = fromRename->size;
858#ifdef DEBUG
859 for (ThreadID tid = 0; tid < numThreads; tid++)
860 assert(insts[tid].empty());
861#endif
862 for (int i = 0; i < insts_from_rename; ++i) {
863 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
864 }
865}
866
867template <class Impl>
868void
869DefaultIEW<Impl>::emptyRenameInsts(ThreadID tid)
870{
871 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions\n", tid);
872
873 while (!insts[tid].empty()) {
874
875 if (insts[tid].front()->isLoad() ||
876 insts[tid].front()->isStore() ) {
877 toRename->iewInfo[tid].dispatchedToLSQ++;
878 }
879
880 toRename->iewInfo[tid].dispatched++;
881
882 insts[tid].pop();
883 }
884}
885
886template <class Impl>
887void
888DefaultIEW<Impl>::wakeCPU()
889{
890 cpu->wakeCPU();
891}
892
893template <class Impl>
894void
895DefaultIEW<Impl>::activityThisCycle()
896{
897 DPRINTF(Activity, "Activity this cycle.\n");
898 cpu->activityThisCycle();
899}
900
901template <class Impl>
902inline void
903DefaultIEW<Impl>::activateStage()
904{
905 DPRINTF(Activity, "Activating stage.\n");
906 cpu->activateStage(O3CPU::IEWIdx);
907}
908
909template <class Impl>
910inline void
911DefaultIEW<Impl>::deactivateStage()
912{
913 DPRINTF(Activity, "Deactivating stage.\n");
914 cpu->deactivateStage(O3CPU::IEWIdx);
915}
916
917template<class Impl>
918void
919DefaultIEW<Impl>::dispatch(ThreadID tid)
920{
921 // If status is Running or idle,
922 // call dispatchInsts()
923 // If status is Unblocking,
924 // buffer any instructions coming from rename
925 // continue trying to empty skid buffer
926 // check if stall conditions have passed
927
928 if (dispatchStatus[tid] == Blocked) {
929 ++iewBlockCycles;
930
931 } else if (dispatchStatus[tid] == Squashing) {
932 ++iewSquashCycles;
933 }
934
935 // Dispatch should try to dispatch as many instructions as its bandwidth
936 // will allow, as long as it is not currently blocked.
937 if (dispatchStatus[tid] == Running ||
938 dispatchStatus[tid] == Idle) {
939 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
940 "dispatch.\n", tid);
941
942 dispatchInsts(tid);
943 } else if (dispatchStatus[tid] == Unblocking) {
944 // Make sure that the skid buffer has something in it if the
945 // status is unblocking.
946 assert(!skidsEmpty());
947
948 // If the status was unblocking, then instructions from the skid
949 // buffer were used. Remove those instructions and handle
950 // the rest of unblocking.
951 dispatchInsts(tid);
952
953 ++iewUnblockCycles;
954
955 if (validInstsFromRename()) {
956 // Add the current inputs to the skid buffer so they can be
957 // reprocessed when this stage unblocks.
958 skidInsert(tid);
959 }
960
961 unblock(tid);
962 }
963}
964
965template <class Impl>
966void
967DefaultIEW<Impl>::dispatchInsts(ThreadID tid)
968{
969 // Obtain instructions from skid buffer if unblocking, or queue from rename
970 // otherwise.
971 std::queue<DynInstPtr> &insts_to_dispatch =
972 dispatchStatus[tid] == Unblocking ?
973 skidBuffer[tid] : insts[tid];
974
975 int insts_to_add = insts_to_dispatch.size();
976
977 DynInstPtr inst;
978 bool add_to_iq = false;
979 int dis_num_inst = 0;
980
981 // Loop through the instructions, putting them in the instruction
982 // queue.
983 for ( ; dis_num_inst < insts_to_add &&
984 dis_num_inst < dispatchWidth;
985 ++dis_num_inst)
986 {
987 inst = insts_to_dispatch.front();
988
989 if (dispatchStatus[tid] == Unblocking) {
990 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
991 "buffer\n", tid);
992 }
993
994 // Make sure there's a valid instruction there.
995 assert(inst);
996
997 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %s [sn:%lli] [tid:%i] to "
998 "IQ.\n",
999 tid, inst->pcState(), inst->seqNum, inst->threadNumber);
1000
1001 // Be sure to mark these instructions as ready so that the
1002 // commit stage can go ahead and execute them, and mark
1003 // them as issued so the IQ doesn't reprocess them.
1004
1005 // Check for squashed instructions.
1006 if (inst->isSquashed()) {
1007 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
1008 "not adding to IQ.\n", tid);
1009
1010 ++iewDispSquashedInsts;
1011
1012 insts_to_dispatch.pop();
1013
1014 //Tell Rename That An Instruction has been processed
1015 if (inst->isLoad() || inst->isStore()) {
1016 toRename->iewInfo[tid].dispatchedToLSQ++;
1017 }
1018 toRename->iewInfo[tid].dispatched++;
1019
1020 continue;
1021 }
1022
1023 // Check for full conditions.
1024 if (instQueue.isFull(tid)) {
1025 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1026
1027 // Call function to start blocking.
1028 block(tid);
1029
1030 // Set unblock to false. Special case where we are using
1031 // skidbuffer (unblocking) instructions but then we still
1032 // get full in the IQ.
1033 toRename->iewUnblock[tid] = false;
1034
1035 ++iewIQFullEvents;
1036 break;
1037 } else if (ldstQueue.isFull(tid)) {
1038 DPRINTF(IEW, "[tid:%i]: Issue: LSQ has become full.\n",tid);
1039
1040 // Call function to start blocking.
1041 block(tid);
1042
1043 // Set unblock to false. Special case where we are using
1044 // skidbuffer (unblocking) instructions but then we still
1045 // get full in the IQ.
1046 toRename->iewUnblock[tid] = false;
1047
1048 ++iewLSQFullEvents;
1049 break;
1050 }
1051
1052 // Otherwise issue the instruction just fine.
1053 if (inst->isLoad()) {
1054 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1055 "encountered, adding to LSQ.\n", tid);
1056
1057 // Reserve a spot in the load store queue for this
1058 // memory access.
1059 ldstQueue.insertLoad(inst);
1060
1061 ++iewDispLoadInsts;
1062
1063 add_to_iq = true;
1064
1065 toRename->iewInfo[tid].dispatchedToLSQ++;
1066 } else if (inst->isStore()) {
1067 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1068 "encountered, adding to LSQ.\n", tid);
1069
1070 ldstQueue.insertStore(inst);
1071
1072 ++iewDispStoreInsts;
1073
1074 if (inst->isStoreConditional()) {
1075 // Store conditionals need to be set as "canCommit()"
1076 // so that commit can process them when they reach the
1077 // head of commit.
1078 // @todo: This is somewhat specific to Alpha.
1079 inst->setCanCommit();
1080 instQueue.insertNonSpec(inst);
1081 add_to_iq = false;
1082
1083 ++iewDispNonSpecInsts;
1084 } else {
1085 add_to_iq = true;
1086 }
1087
1088 toRename->iewInfo[tid].dispatchedToLSQ++;
1089 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1090 // Same as non-speculative stores.
1091 inst->setCanCommit();
1092 instQueue.insertBarrier(inst);
1093 add_to_iq = false;
1094 } else if (inst->isNop()) {
1095 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1096 "skipping.\n", tid);
1097
1098 inst->setIssued();
1099 inst->setExecuted();
1100 inst->setCanCommit();
1101
1102 instQueue.recordProducer(inst);
1103
1104 iewExecutedNop[tid]++;
1105
1106 add_to_iq = false;
1107 } else if (inst->isExecuted()) {
1108 assert(0 && "Instruction shouldn't be executed.\n");
1109 DPRINTF(IEW, "Issue: Executed branch encountered, "
1110 "skipping.\n");
1111
1112 inst->setIssued();
1113 inst->setCanCommit();
1114
1115 instQueue.recordProducer(inst);
1116
1117 add_to_iq = false;
1118 } else {
1119 add_to_iq = true;
1120 }
1121 if (inst->isNonSpeculative()) {
1122 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1123 "encountered, skipping.\n", tid);
1124
1125 // Same as non-speculative stores.
1126 inst->setCanCommit();
1127
1128 // Specifically insert it as nonspeculative.
1129 instQueue.insertNonSpec(inst);
1130
1131 ++iewDispNonSpecInsts;
1132
1133 add_to_iq = false;
1134 }
1135
1136 // If the instruction queue is not full, then add the
1137 // instruction.
1138 if (add_to_iq) {
1139 instQueue.insert(inst);
1140 }
1141
1142 insts_to_dispatch.pop();
1143
1144 toRename->iewInfo[tid].dispatched++;
1145
1146 ++iewDispatchedInsts;
1147
1148#if TRACING_ON
1149 inst->dispatchTick = curTick() - inst->fetchTick;
1150#endif
1151 }
1152
1153 if (!insts_to_dispatch.empty()) {
1154 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1155 block(tid);
1156 toRename->iewUnblock[tid] = false;
1157 }
1158
1159 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1160 dispatchStatus[tid] = Running;
1161
1162 updatedQueues = true;
1163 }
1164
1165 dis_num_inst = 0;
1166}
1167
1168template <class Impl>
1169void
1170DefaultIEW<Impl>::printAvailableInsts()
1171{
1172 int inst = 0;
1173
1174 std::cout << "Available Instructions: ";
1175
1176 while (fromIssue->insts[inst]) {
1177
1178 if (inst%3==0) std::cout << "\n\t";
1179
1180 std::cout << "PC: " << fromIssue->insts[inst]->pcState()
1181 << " TN: " << fromIssue->insts[inst]->threadNumber
1182 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1183
1184 inst++;
1185
1186 }
1187
1188 std::cout << "\n";
1189}
1190
1191template <class Impl>
1192void
1193DefaultIEW<Impl>::executeInsts()
1194{
1195 wbNumInst = 0;
1196 wbCycle = 0;
1197
1198 list<ThreadID>::iterator threads = activeThreads->begin();
1199 list<ThreadID>::iterator end = activeThreads->end();
1200
1201 while (threads != end) {
1202 ThreadID tid = *threads++;
1203 fetchRedirect[tid] = false;
1204 }
1205
1206 // Uncomment this if you want to see all available instructions.
1207 // @todo This doesn't actually work anymore, we should fix it.
1208// printAvailableInsts();
1209
1210 // Execute/writeback any instructions that are available.
1211 int insts_to_execute = fromIssue->size;
1212 int inst_num = 0;
1213 for (; inst_num < insts_to_execute;
1214 ++inst_num) {
1215
1216 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1217
1218 DynInstPtr inst = instQueue.getInstToExecute();
1219
1220 DPRINTF(IEW, "Execute: Processing PC %s, [tid:%i] [sn:%i].\n",
1221 inst->pcState(), inst->threadNumber,inst->seqNum);
1222
1223 // Check if the instruction is squashed; if so then skip it
1224 if (inst->isSquashed()) {
1225 DPRINTF(IEW, "Execute: Instruction was squashed. PC: %s, [tid:%i]"
1226 " [sn:%i]\n", inst->pcState(), inst->threadNumber,
1227 inst->seqNum);
1228
1229 // Consider this instruction executed so that commit can go
1230 // ahead and retire the instruction.
1231 inst->setExecuted();
1232
1233 // Not sure if I should set this here or just let commit try to
1234 // commit any squashed instructions. I like the latter a bit more.
1235 inst->setCanCommit();
1236
1237 ++iewExecSquashedInsts;
1238
1239 decrWb(inst->seqNum);
1240 continue;
1241 }
1242
1243 Fault fault = NoFault;
1244
1245 // Execute instruction.
1246 // Note that if the instruction faults, it will be handled
1247 // at the commit stage.
1248 if (inst->isMemRef()) {
1249 DPRINTF(IEW, "Execute: Calculating address for memory "
1250 "reference.\n");
1251
1252 // Tell the LDSTQ to execute this instruction (if it is a load).
1253 if (inst->isLoad()) {
1254 // Loads will mark themselves as executed, and their writeback
1255 // event adds the instruction to the queue to commit
1256 fault = ldstQueue.executeLoad(inst);
1257
1258 if (inst->isTranslationDelayed() &&
1259 fault == NoFault) {
1260 // A hw page table walk is currently going on; the
1261 // instruction must be deferred.
1262 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1263 "load.\n");
1264 instQueue.deferMemInst(inst);
1265 continue;
1266 }
1267
1268 if (inst->isDataPrefetch() || inst->isInstPrefetch()) {
1269 inst->fault = NoFault;
1270 }
1271 } else if (inst->isStore()) {
1272 fault = ldstQueue.executeStore(inst);
1273
1274 if (inst->isTranslationDelayed() &&
1275 fault == NoFault) {
1276 // A hw page table walk is currently going on; the
1277 // instruction must be deferred.
1278 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1279 "store.\n");
1280 instQueue.deferMemInst(inst);
1281 continue;
1282 }
1283
1284 // If the store had a fault then it may not have a mem req
1285 if (fault != NoFault || inst->readPredicate() == false ||
1286 !inst->isStoreConditional()) {
1287 // If the instruction faulted, then we need to send it along
1288 // to commit without the instruction completing.
1289 // Send this instruction to commit, also make sure iew stage
1290 // realizes there is activity.
1291 inst->setExecuted();
1292 instToCommit(inst);
1293 activityThisCycle();
1294 }
1295
1296 // Store conditionals will mark themselves as
1297 // executed, and their writeback event will add the
1298 // instruction to the queue to commit.
1299 } else {
1300 panic("Unexpected memory type!\n");
1301 }
1302
1303 } else {
1304 // If the instruction has already faulted, then skip executing it.
1305 // Such case can happen when it faulted during ITLB translation.
1306 // If we execute the instruction (even if it's a nop) the fault
1307 // will be replaced and we will lose it.
1308 if (inst->getFault() == NoFault) {
1309 inst->execute();
1310 if (inst->readPredicate() == false)
1311 inst->forwardOldRegs();
1312 }
1313
1314 inst->setExecuted();
1315
1316 instToCommit(inst);
1317 }
1318
1319 updateExeInstStats(inst);
1320
1321 // Check if branch prediction was correct, if not then we need
1322 // to tell commit to squash in flight instructions. Only
1323 // handle this if there hasn't already been something that
1324 // redirects fetch in this group of instructions.
1325
1326 // This probably needs to prioritize the redirects if a different
1327 // scheduler is used. Currently the scheduler schedules the oldest
1328 // instruction first, so the branch resolution order will be correct.
1329 ThreadID tid = inst->threadNumber;
1330
1331 if (!fetchRedirect[tid] ||
1332 !toCommit->squash[tid] ||
1333 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1334
1335 // Prevent testing for misprediction on load instructions,
1336 // that have not been executed.
1337 bool loadNotExecuted = !inst->isExecuted() && inst->isLoad();
1338
1339 if (inst->mispredicted() && !loadNotExecuted) {
1340 fetchRedirect[tid] = true;
1341
1342 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1343 DPRINTF(IEW, "Predicted target was PC: %s.\n",
1344 inst->readPredTarg());
1345 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %s.\n",
1346 inst->pcState());
1347 // If incorrect, then signal the ROB that it must be squashed.
1348 squashDueToBranch(inst, tid);
1349
1350 if (inst->readPredTaken()) {
1351 predictedTakenIncorrect++;
1352 } else {
1353 predictedNotTakenIncorrect++;
1354 }
1355 } else if (ldstQueue.violation(tid)) {
1356 assert(inst->isMemRef());
1357 // If there was an ordering violation, then get the
1358 // DynInst that caused the violation. Note that this
1359 // clears the violation signal.
1360 DynInstPtr violator;
1361 violator = ldstQueue.getMemDepViolator(tid);
1362
1363 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: %s "
1364 "[sn:%lli], inst PC: %s [sn:%lli]. Addr is: %#x.\n",
1365 violator->pcState(), violator->seqNum,
1366 inst->pcState(), inst->seqNum, inst->physEffAddr);
1367
1368 fetchRedirect[tid] = true;
1369
1370 // Tell the instruction queue that a violation has occured.
1371 instQueue.violation(inst, violator);
1372
1373 // Squash.
1374 squashDueToMemOrder(violator, tid);
1375
1376 ++memOrderViolationEvents;
1377 } else if (ldstQueue.loadBlocked(tid) &&
1378 !ldstQueue.isLoadBlockedHandled(tid)) {
1379 fetchRedirect[tid] = true;
1380
1381 DPRINTF(IEW, "Load operation couldn't execute because the "
1382 "memory system is blocked. PC: %s [sn:%lli]\n",
1383 inst->pcState(), inst->seqNum);
1384
1385 squashDueToMemBlocked(inst, tid);
1386 }
1387 } else {
1388 // Reset any state associated with redirects that will not
1389 // be used.
1390 if (ldstQueue.violation(tid)) {
1391 assert(inst->isMemRef());
1392
1393 DynInstPtr violator = ldstQueue.getMemDepViolator(tid);
1394
1395 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1396 "%s, inst PC: %s. Addr is: %#x.\n",
1397 violator->pcState(), inst->pcState(),
1398 inst->physEffAddr);
1399 DPRINTF(IEW, "Violation will not be handled because "
1400 "already squashing\n");
1401
1402 ++memOrderViolationEvents;
1403 }
1404 if (ldstQueue.loadBlocked(tid) &&
1405 !ldstQueue.isLoadBlockedHandled(tid)) {
1406 DPRINTF(IEW, "Load operation couldn't execute because the "
1407 "memory system is blocked. PC: %s [sn:%lli]\n",
1408 inst->pcState(), inst->seqNum);
1409 DPRINTF(IEW, "Blocked load will not be handled because "
1410 "already squashing\n");
1411
1412 ldstQueue.setLoadBlockedHandled(tid);
1413 }
1414
1415 }
1416 }
1417
1418 // Update and record activity if we processed any instructions.
1419 if (inst_num) {
1420 if (exeStatus == Idle) {
1421 exeStatus = Running;
1422 }
1423
1424 updatedQueues = true;
1425
1426 cpu->activityThisCycle();
1427 }
1428
1429 // Need to reset this in case a writeback event needs to write into the
1430 // iew queue. That way the writeback event will write into the correct
1431 // spot in the queue.
1432 wbNumInst = 0;
1433
1434}
1435
1436template <class Impl>
1437void
1438DefaultIEW<Impl>::writebackInsts()
1439{
1440 // Loop through the head of the time buffer and wake any
1441 // dependents. These instructions are about to write back. Also
1442 // mark scoreboard that this instruction is finally complete.
1443 // Either have IEW have direct access to scoreboard, or have this
1444 // as part of backwards communication.
1445 for (int inst_num = 0; inst_num < wbWidth &&
1446 toCommit->insts[inst_num]; inst_num++) {
1447 DynInstPtr inst = toCommit->insts[inst_num];
1448 ThreadID tid = inst->threadNumber;
1449
1450 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %s.\n",
1451 inst->seqNum, inst->pcState());
1452
1453 iewInstsToCommit[tid]++;
1454
1455 // Some instructions will be sent to commit without having
1456 // executed because they need commit to handle them.
1457 // E.g. Uncached loads have not actually executed when they
1458 // are first sent to commit. Instead commit must tell the LSQ
1459 // when it's ready to execute the uncached load.
1460 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1461 int dependents = instQueue.wakeDependents(inst);
1462
1463 for (int i = 0; i < inst->numDestRegs(); i++) {
1464 //mark as Ready
1465 DPRINTF(IEW,"Setting Destination Register %i\n",
1466 inst->renamedDestRegIdx(i));
1467 scoreboard->setReg(inst->renamedDestRegIdx(i));
1468 }
1469
1470 if (dependents) {
1471 producerInst[tid]++;
1472 consumerInst[tid]+= dependents;
1473 }
1474 writebackCount[tid]++;
1475 }
1476
1477 decrWb(inst->seqNum);
1478 }
1479}
1480
1481template<class Impl>
1482void
1483DefaultIEW<Impl>::tick()
1484{
1485 wbNumInst = 0;
1486 wbCycle = 0;
1487
1488 wroteToTimeBuffer = false;
1489 updatedQueues = false;
1490
1491 sortInsts();
1492
1493 // Free function units marked as being freed this cycle.
1494 fuPool->processFreeUnits();
1495
1496 list<ThreadID>::iterator threads = activeThreads->begin();
1497 list<ThreadID>::iterator end = activeThreads->end();
1498
1499 // Check stall and squash signals, dispatch any instructions.
1500 while (threads != end) {
1501 ThreadID tid = *threads++;
1502
1503 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1504
1505 checkSignalsAndUpdate(tid);
1506 dispatch(tid);
1507 }
1508
1509 if (exeStatus != Squashing) {
1510 executeInsts();
1511
1512 writebackInsts();
1513
1514 // Have the instruction queue try to schedule any ready instructions.
1515 // (In actuality, this scheduling is for instructions that will
1516 // be executed next cycle.)
1517 instQueue.scheduleReadyInsts();
1518
1519 // Also should advance its own time buffers if the stage ran.
1520 // Not the best place for it, but this works (hopefully).
1521 issueToExecQueue.advance();
1522 }
1523
1524 bool broadcast_free_entries = false;
1525
1526 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1527 exeStatus = Idle;
1528 updateLSQNextCycle = false;
1529
1530 broadcast_free_entries = true;
1531 }
1532
1533 // Writeback any stores using any leftover bandwidth.
1534 ldstQueue.writebackStores();
1535
1536 // Check the committed load/store signals to see if there's a load
1537 // or store to commit. Also check if it's being told to execute a
1538 // nonspeculative instruction.
1539 // This is pretty inefficient...
1540
1541 threads = activeThreads->begin();
1542 while (threads != end) {
1543 ThreadID tid = (*threads++);
1544
1545 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1546
1547 // Update structures based on instructions committed.
1548 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1549 !fromCommit->commitInfo[tid].squash &&
1550 !fromCommit->commitInfo[tid].robSquashing) {
1551
1552 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1553
1554 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1555
1556 updateLSQNextCycle = true;
1557 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1558 }
1559
1560 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1561
1562 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1563 if (fromCommit->commitInfo[tid].uncached) {
1564 instQueue.replayMemInst(fromCommit->commitInfo[tid].uncachedLoad);
1565 fromCommit->commitInfo[tid].uncachedLoad->setAtCommit();
1566 } else {
1567 instQueue.scheduleNonSpec(
1568 fromCommit->commitInfo[tid].nonSpecSeqNum);
1569 }
1570 }
1571
1572 if (broadcast_free_entries) {
1573 toFetch->iewInfo[tid].iqCount =
1574 instQueue.getCount(tid);
1575 toFetch->iewInfo[tid].ldstqCount =
1576 ldstQueue.getCount(tid);
1577
1578 toRename->iewInfo[tid].usedIQ = true;
1579 toRename->iewInfo[tid].freeIQEntries =
1580 instQueue.numFreeEntries();
1581 toRename->iewInfo[tid].usedLSQ = true;
1582 toRename->iewInfo[tid].freeLSQEntries =
1583 ldstQueue.numFreeEntries(tid);
1584
1585 wroteToTimeBuffer = true;
1586 }
1587
1588 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1589 tid, toRename->iewInfo[tid].dispatched);
1590 }
1591
1592 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1593 "LSQ has %i free entries.\n",
1594 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1595 ldstQueue.numFreeEntries());
1596
1597 updateStatus();
1598
1599 if (wroteToTimeBuffer) {
1600 DPRINTF(Activity, "Activity this cycle.\n");
1601 cpu->activityThisCycle();
1602 }
1603}
1604
1605template <class Impl>
1606void
1607DefaultIEW<Impl>::updateExeInstStats(DynInstPtr &inst)
1608{
1609 ThreadID tid = inst->threadNumber;
1610
1611 iewExecutedInsts++;
1612
1613#if TRACING_ON
60#include "params/DerivO3CPU.hh"
61
62using namespace std;
63
64template<class Impl>
65DefaultIEW<Impl>::DefaultIEW(O3CPU *_cpu, DerivO3CPUParams *params)
66 : issueToExecQueue(params->backComSize, params->forwardComSize),
67 cpu(_cpu),
68 instQueue(_cpu, this, params),
69 ldstQueue(_cpu, this, params),
70 fuPool(params->fuPool),
71 commitToIEWDelay(params->commitToIEWDelay),
72 renameToIEWDelay(params->renameToIEWDelay),
73 issueToExecuteDelay(params->issueToExecuteDelay),
74 dispatchWidth(params->dispatchWidth),
75 issueWidth(params->issueWidth),
76 wbOutstanding(0),
77 wbWidth(params->wbWidth),
78 numThreads(params->numThreads)
79{
80 _status = Active;
81 exeStatus = Running;
82 wbStatus = Idle;
83
84 // Setup wire to read instructions coming from issue.
85 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
86
87 // Instruction queue needs the queue between issue and execute.
88 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
89
90 for (ThreadID tid = 0; tid < numThreads; tid++) {
91 dispatchStatus[tid] = Running;
92 stalls[tid].commit = false;
93 fetchRedirect[tid] = false;
94 }
95
96 wbMax = wbWidth * params->wbDepth;
97
98 updateLSQNextCycle = false;
99
100 ableToIssue = true;
101
102 skidBufferMax = (3 * (renameToIEWDelay * params->renameWidth)) + issueWidth;
103}
104
105template <class Impl>
106std::string
107DefaultIEW<Impl>::name() const
108{
109 return cpu->name() + ".iew";
110}
111
112template <class Impl>
113void
114DefaultIEW<Impl>::regStats()
115{
116 using namespace Stats;
117
118 instQueue.regStats();
119 ldstQueue.regStats();
120
121 iewIdleCycles
122 .name(name() + ".iewIdleCycles")
123 .desc("Number of cycles IEW is idle");
124
125 iewSquashCycles
126 .name(name() + ".iewSquashCycles")
127 .desc("Number of cycles IEW is squashing");
128
129 iewBlockCycles
130 .name(name() + ".iewBlockCycles")
131 .desc("Number of cycles IEW is blocking");
132
133 iewUnblockCycles
134 .name(name() + ".iewUnblockCycles")
135 .desc("Number of cycles IEW is unblocking");
136
137 iewDispatchedInsts
138 .name(name() + ".iewDispatchedInsts")
139 .desc("Number of instructions dispatched to IQ");
140
141 iewDispSquashedInsts
142 .name(name() + ".iewDispSquashedInsts")
143 .desc("Number of squashed instructions skipped by dispatch");
144
145 iewDispLoadInsts
146 .name(name() + ".iewDispLoadInsts")
147 .desc("Number of dispatched load instructions");
148
149 iewDispStoreInsts
150 .name(name() + ".iewDispStoreInsts")
151 .desc("Number of dispatched store instructions");
152
153 iewDispNonSpecInsts
154 .name(name() + ".iewDispNonSpecInsts")
155 .desc("Number of dispatched non-speculative instructions");
156
157 iewIQFullEvents
158 .name(name() + ".iewIQFullEvents")
159 .desc("Number of times the IQ has become full, causing a stall");
160
161 iewLSQFullEvents
162 .name(name() + ".iewLSQFullEvents")
163 .desc("Number of times the LSQ has become full, causing a stall");
164
165 memOrderViolationEvents
166 .name(name() + ".memOrderViolationEvents")
167 .desc("Number of memory order violations");
168
169 predictedTakenIncorrect
170 .name(name() + ".predictedTakenIncorrect")
171 .desc("Number of branches that were predicted taken incorrectly");
172
173 predictedNotTakenIncorrect
174 .name(name() + ".predictedNotTakenIncorrect")
175 .desc("Number of branches that were predicted not taken incorrectly");
176
177 branchMispredicts
178 .name(name() + ".branchMispredicts")
179 .desc("Number of branch mispredicts detected at execute");
180
181 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
182
183 iewExecutedInsts
184 .name(name() + ".iewExecutedInsts")
185 .desc("Number of executed instructions");
186
187 iewExecLoadInsts
188 .init(cpu->numThreads)
189 .name(name() + ".iewExecLoadInsts")
190 .desc("Number of load instructions executed")
191 .flags(total);
192
193 iewExecSquashedInsts
194 .name(name() + ".iewExecSquashedInsts")
195 .desc("Number of squashed instructions skipped in execute");
196
197 iewExecutedSwp
198 .init(cpu->numThreads)
199 .name(name() + ".exec_swp")
200 .desc("number of swp insts executed")
201 .flags(total);
202
203 iewExecutedNop
204 .init(cpu->numThreads)
205 .name(name() + ".exec_nop")
206 .desc("number of nop insts executed")
207 .flags(total);
208
209 iewExecutedRefs
210 .init(cpu->numThreads)
211 .name(name() + ".exec_refs")
212 .desc("number of memory reference insts executed")
213 .flags(total);
214
215 iewExecutedBranches
216 .init(cpu->numThreads)
217 .name(name() + ".exec_branches")
218 .desc("Number of branches executed")
219 .flags(total);
220
221 iewExecStoreInsts
222 .name(name() + ".exec_stores")
223 .desc("Number of stores executed")
224 .flags(total);
225 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
226
227 iewExecRate
228 .name(name() + ".exec_rate")
229 .desc("Inst execution rate")
230 .flags(total);
231
232 iewExecRate = iewExecutedInsts / cpu->numCycles;
233
234 iewInstsToCommit
235 .init(cpu->numThreads)
236 .name(name() + ".wb_sent")
237 .desc("cumulative count of insts sent to commit")
238 .flags(total);
239
240 writebackCount
241 .init(cpu->numThreads)
242 .name(name() + ".wb_count")
243 .desc("cumulative count of insts written-back")
244 .flags(total);
245
246 producerInst
247 .init(cpu->numThreads)
248 .name(name() + ".wb_producers")
249 .desc("num instructions producing a value")
250 .flags(total);
251
252 consumerInst
253 .init(cpu->numThreads)
254 .name(name() + ".wb_consumers")
255 .desc("num instructions consuming a value")
256 .flags(total);
257
258 wbPenalized
259 .init(cpu->numThreads)
260 .name(name() + ".wb_penalized")
261 .desc("number of instrctions required to write to 'other' IQ")
262 .flags(total);
263
264 wbPenalizedRate
265 .name(name() + ".wb_penalized_rate")
266 .desc ("fraction of instructions written-back that wrote to 'other' IQ")
267 .flags(total);
268
269 wbPenalizedRate = wbPenalized / writebackCount;
270
271 wbFanout
272 .name(name() + ".wb_fanout")
273 .desc("average fanout of values written-back")
274 .flags(total);
275
276 wbFanout = producerInst / consumerInst;
277
278 wbRate
279 .name(name() + ".wb_rate")
280 .desc("insts written-back per cycle")
281 .flags(total);
282 wbRate = writebackCount / cpu->numCycles;
283}
284
285template<class Impl>
286void
287DefaultIEW<Impl>::startupStage()
288{
289 for (ThreadID tid = 0; tid < numThreads; tid++) {
290 toRename->iewInfo[tid].usedIQ = true;
291 toRename->iewInfo[tid].freeIQEntries =
292 instQueue.numFreeEntries(tid);
293
294 toRename->iewInfo[tid].usedLSQ = true;
295 toRename->iewInfo[tid].freeLSQEntries =
296 ldstQueue.numFreeEntries(tid);
297 }
298
299 // Initialize the checker's dcache port here
300 if (cpu->checker) {
301 cpu->checker->setDcachePort(&cpu->getDataPort());
302 }
303
304 cpu->activateStage(O3CPU::IEWIdx);
305}
306
307template<class Impl>
308void
309DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
310{
311 timeBuffer = tb_ptr;
312
313 // Setup wire to read information from time buffer, from commit.
314 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
315
316 // Setup wire to write information back to previous stages.
317 toRename = timeBuffer->getWire(0);
318
319 toFetch = timeBuffer->getWire(0);
320
321 // Instruction queue also needs main time buffer.
322 instQueue.setTimeBuffer(tb_ptr);
323}
324
325template<class Impl>
326void
327DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
328{
329 renameQueue = rq_ptr;
330
331 // Setup wire to read information from rename queue.
332 fromRename = renameQueue->getWire(-renameToIEWDelay);
333}
334
335template<class Impl>
336void
337DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
338{
339 iewQueue = iq_ptr;
340
341 // Setup wire to write instructions to commit.
342 toCommit = iewQueue->getWire(0);
343}
344
345template<class Impl>
346void
347DefaultIEW<Impl>::setActiveThreads(list<ThreadID> *at_ptr)
348{
349 activeThreads = at_ptr;
350
351 ldstQueue.setActiveThreads(at_ptr);
352 instQueue.setActiveThreads(at_ptr);
353}
354
355template<class Impl>
356void
357DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
358{
359 scoreboard = sb_ptr;
360}
361
362template <class Impl>
363bool
364DefaultIEW<Impl>::isDrained() const
365{
366 bool drained(ldstQueue.isDrained());
367
368 for (ThreadID tid = 0; tid < numThreads; tid++) {
369 if (!insts[tid].empty()) {
370 DPRINTF(Drain, "%i: Insts not empty.\n", tid);
371 drained = false;
372 }
373 if (!skidBuffer[tid].empty()) {
374 DPRINTF(Drain, "%i: Skid buffer not empty.\n", tid);
375 drained = false;
376 }
377 }
378
379 return drained;
380}
381
382template <class Impl>
383void
384DefaultIEW<Impl>::drainSanityCheck() const
385{
386 assert(isDrained());
387
388 instQueue.drainSanityCheck();
389 ldstQueue.drainSanityCheck();
390 fuPool->drainSanityCheck();
391}
392
393template <class Impl>
394void
395DefaultIEW<Impl>::takeOverFrom()
396{
397 // Reset all state.
398 _status = Active;
399 exeStatus = Running;
400 wbStatus = Idle;
401
402 instQueue.takeOverFrom();
403 ldstQueue.takeOverFrom();
404 fuPool->takeOverFrom();
405
406 startupStage();
407 cpu->activityThisCycle();
408
409 for (ThreadID tid = 0; tid < numThreads; tid++) {
410 dispatchStatus[tid] = Running;
411 stalls[tid].commit = false;
412 fetchRedirect[tid] = false;
413 }
414
415 updateLSQNextCycle = false;
416
417 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
418 issueToExecQueue.advance();
419 }
420}
421
422template<class Impl>
423void
424DefaultIEW<Impl>::squash(ThreadID tid)
425{
426 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n", tid);
427
428 // Tell the IQ to start squashing.
429 instQueue.squash(tid);
430
431 // Tell the LDSTQ to start squashing.
432 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
433 updatedQueues = true;
434
435 // Clear the skid buffer in case it has any data in it.
436 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
437 tid, fromCommit->commitInfo[tid].doneSeqNum);
438
439 while (!skidBuffer[tid].empty()) {
440 if (skidBuffer[tid].front()->isLoad() ||
441 skidBuffer[tid].front()->isStore() ) {
442 toRename->iewInfo[tid].dispatchedToLSQ++;
443 }
444
445 toRename->iewInfo[tid].dispatched++;
446
447 skidBuffer[tid].pop();
448 }
449
450 emptyRenameInsts(tid);
451}
452
453template<class Impl>
454void
455DefaultIEW<Impl>::squashDueToBranch(DynInstPtr &inst, ThreadID tid)
456{
457 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %s "
458 "[sn:%i].\n", tid, inst->pcState(), inst->seqNum);
459
460 if (toCommit->squash[tid] == false ||
461 inst->seqNum < toCommit->squashedSeqNum[tid]) {
462 toCommit->squash[tid] = true;
463 toCommit->squashedSeqNum[tid] = inst->seqNum;
464 toCommit->branchTaken[tid] = inst->pcState().branching();
465
466 TheISA::PCState pc = inst->pcState();
467 TheISA::advancePC(pc, inst->staticInst);
468
469 toCommit->pc[tid] = pc;
470 toCommit->mispredictInst[tid] = inst;
471 toCommit->includeSquashInst[tid] = false;
472
473 wroteToTimeBuffer = true;
474 }
475
476}
477
478template<class Impl>
479void
480DefaultIEW<Impl>::squashDueToMemOrder(DynInstPtr &inst, ThreadID tid)
481{
482 DPRINTF(IEW, "[tid:%i]: Memory violation, squashing violator and younger "
483 "insts, PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
484 // Need to include inst->seqNum in the following comparison to cover the
485 // corner case when a branch misprediction and a memory violation for the
486 // same instruction (e.g. load PC) are detected in the same cycle. In this
487 // case the memory violator should take precedence over the branch
488 // misprediction because it requires the violator itself to be included in
489 // the squash.
490 if (toCommit->squash[tid] == false ||
491 inst->seqNum <= toCommit->squashedSeqNum[tid]) {
492 toCommit->squash[tid] = true;
493
494 toCommit->squashedSeqNum[tid] = inst->seqNum;
495 toCommit->pc[tid] = inst->pcState();
496 toCommit->mispredictInst[tid] = NULL;
497
498 // Must include the memory violator in the squash.
499 toCommit->includeSquashInst[tid] = true;
500
501 wroteToTimeBuffer = true;
502 }
503}
504
505template<class Impl>
506void
507DefaultIEW<Impl>::squashDueToMemBlocked(DynInstPtr &inst, ThreadID tid)
508{
509 DPRINTF(IEW, "[tid:%i]: Memory blocked, squashing load and younger insts, "
510 "PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
511 if (toCommit->squash[tid] == false ||
512 inst->seqNum < toCommit->squashedSeqNum[tid]) {
513 toCommit->squash[tid] = true;
514
515 toCommit->squashedSeqNum[tid] = inst->seqNum;
516 toCommit->pc[tid] = inst->pcState();
517 toCommit->mispredictInst[tid] = NULL;
518
519 // Must include the broadcasted SN in the squash.
520 toCommit->includeSquashInst[tid] = true;
521
522 ldstQueue.setLoadBlockedHandled(tid);
523
524 wroteToTimeBuffer = true;
525 }
526}
527
528template<class Impl>
529void
530DefaultIEW<Impl>::block(ThreadID tid)
531{
532 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
533
534 if (dispatchStatus[tid] != Blocked &&
535 dispatchStatus[tid] != Unblocking) {
536 toRename->iewBlock[tid] = true;
537 wroteToTimeBuffer = true;
538 }
539
540 // Add the current inputs to the skid buffer so they can be
541 // reprocessed when this stage unblocks.
542 skidInsert(tid);
543
544 dispatchStatus[tid] = Blocked;
545}
546
547template<class Impl>
548void
549DefaultIEW<Impl>::unblock(ThreadID tid)
550{
551 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
552 "buffer %u.\n",tid, tid);
553
554 // If the skid bufffer is empty, signal back to previous stages to unblock.
555 // Also switch status to running.
556 if (skidBuffer[tid].empty()) {
557 toRename->iewUnblock[tid] = true;
558 wroteToTimeBuffer = true;
559 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
560 dispatchStatus[tid] = Running;
561 }
562}
563
564template<class Impl>
565void
566DefaultIEW<Impl>::wakeDependents(DynInstPtr &inst)
567{
568 instQueue.wakeDependents(inst);
569}
570
571template<class Impl>
572void
573DefaultIEW<Impl>::rescheduleMemInst(DynInstPtr &inst)
574{
575 instQueue.rescheduleMemInst(inst);
576}
577
578template<class Impl>
579void
580DefaultIEW<Impl>::replayMemInst(DynInstPtr &inst)
581{
582 instQueue.replayMemInst(inst);
583}
584
585template<class Impl>
586void
587DefaultIEW<Impl>::instToCommit(DynInstPtr &inst)
588{
589 // This function should not be called after writebackInsts in a
590 // single cycle. That will cause problems with an instruction
591 // being added to the queue to commit without being processed by
592 // writebackInsts prior to being sent to commit.
593
594 // First check the time slot that this instruction will write
595 // to. If there are free write ports at the time, then go ahead
596 // and write the instruction to that time. If there are not,
597 // keep looking back to see where's the first time there's a
598 // free slot.
599 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
600 ++wbNumInst;
601 if (wbNumInst == wbWidth) {
602 ++wbCycle;
603 wbNumInst = 0;
604 }
605
606 assert((wbCycle * wbWidth + wbNumInst) <= wbMax);
607 }
608
609 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
610 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
611 // Add finished instruction to queue to commit.
612 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
613 (*iewQueue)[wbCycle].size++;
614}
615
616template <class Impl>
617unsigned
618DefaultIEW<Impl>::validInstsFromRename()
619{
620 unsigned inst_count = 0;
621
622 for (int i=0; i<fromRename->size; i++) {
623 if (!fromRename->insts[i]->isSquashed())
624 inst_count++;
625 }
626
627 return inst_count;
628}
629
630template<class Impl>
631void
632DefaultIEW<Impl>::skidInsert(ThreadID tid)
633{
634 DynInstPtr inst = NULL;
635
636 while (!insts[tid].empty()) {
637 inst = insts[tid].front();
638
639 insts[tid].pop();
640
641 DPRINTF(Decode,"[tid:%i]: Inserting [sn:%lli] PC:%s into "
642 "dispatch skidBuffer %i\n",tid, inst->seqNum,
643 inst->pcState(),tid);
644
645 skidBuffer[tid].push(inst);
646 }
647
648 assert(skidBuffer[tid].size() <= skidBufferMax &&
649 "Skidbuffer Exceeded Max Size");
650}
651
652template<class Impl>
653int
654DefaultIEW<Impl>::skidCount()
655{
656 int max=0;
657
658 list<ThreadID>::iterator threads = activeThreads->begin();
659 list<ThreadID>::iterator end = activeThreads->end();
660
661 while (threads != end) {
662 ThreadID tid = *threads++;
663 unsigned thread_count = skidBuffer[tid].size();
664 if (max < thread_count)
665 max = thread_count;
666 }
667
668 return max;
669}
670
671template<class Impl>
672bool
673DefaultIEW<Impl>::skidsEmpty()
674{
675 list<ThreadID>::iterator threads = activeThreads->begin();
676 list<ThreadID>::iterator end = activeThreads->end();
677
678 while (threads != end) {
679 ThreadID tid = *threads++;
680
681 if (!skidBuffer[tid].empty())
682 return false;
683 }
684
685 return true;
686}
687
688template <class Impl>
689void
690DefaultIEW<Impl>::updateStatus()
691{
692 bool any_unblocking = false;
693
694 list<ThreadID>::iterator threads = activeThreads->begin();
695 list<ThreadID>::iterator end = activeThreads->end();
696
697 while (threads != end) {
698 ThreadID tid = *threads++;
699
700 if (dispatchStatus[tid] == Unblocking) {
701 any_unblocking = true;
702 break;
703 }
704 }
705
706 // If there are no ready instructions waiting to be scheduled by the IQ,
707 // and there's no stores waiting to write back, and dispatch is not
708 // unblocking, then there is no internal activity for the IEW stage.
709 instQueue.intInstQueueReads++;
710 if (_status == Active && !instQueue.hasReadyInsts() &&
711 !ldstQueue.willWB() && !any_unblocking) {
712 DPRINTF(IEW, "IEW switching to idle\n");
713
714 deactivateStage();
715
716 _status = Inactive;
717 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
718 ldstQueue.willWB() ||
719 any_unblocking)) {
720 // Otherwise there is internal activity. Set to active.
721 DPRINTF(IEW, "IEW switching to active\n");
722
723 activateStage();
724
725 _status = Active;
726 }
727}
728
729template <class Impl>
730void
731DefaultIEW<Impl>::resetEntries()
732{
733 instQueue.resetEntries();
734 ldstQueue.resetEntries();
735}
736
737template <class Impl>
738void
739DefaultIEW<Impl>::readStallSignals(ThreadID tid)
740{
741 if (fromCommit->commitBlock[tid]) {
742 stalls[tid].commit = true;
743 }
744
745 if (fromCommit->commitUnblock[tid]) {
746 assert(stalls[tid].commit);
747 stalls[tid].commit = false;
748 }
749}
750
751template <class Impl>
752bool
753DefaultIEW<Impl>::checkStall(ThreadID tid)
754{
755 bool ret_val(false);
756
757 if (stalls[tid].commit) {
758 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
759 ret_val = true;
760 } else if (instQueue.isFull(tid)) {
761 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
762 ret_val = true;
763 } else if (ldstQueue.isFull(tid)) {
764 DPRINTF(IEW,"[tid:%i]: Stall: LSQ is full\n",tid);
765
766 if (ldstQueue.numLoads(tid) > 0 ) {
767
768 DPRINTF(IEW,"[tid:%i]: LSQ oldest load: [sn:%i] \n",
769 tid,ldstQueue.getLoadHeadSeqNum(tid));
770 }
771
772 if (ldstQueue.numStores(tid) > 0) {
773
774 DPRINTF(IEW,"[tid:%i]: LSQ oldest store: [sn:%i] \n",
775 tid,ldstQueue.getStoreHeadSeqNum(tid));
776 }
777
778 ret_val = true;
779 } else if (ldstQueue.isStalled(tid)) {
780 DPRINTF(IEW,"[tid:%i]: Stall: LSQ stall detected.\n",tid);
781 ret_val = true;
782 }
783
784 return ret_val;
785}
786
787template <class Impl>
788void
789DefaultIEW<Impl>::checkSignalsAndUpdate(ThreadID tid)
790{
791 // Check if there's a squash signal, squash if there is
792 // Check stall signals, block if there is.
793 // If status was Blocked
794 // if so then go to unblocking
795 // If status was Squashing
796 // check if squashing is not high. Switch to running this cycle.
797
798 readStallSignals(tid);
799
800 if (fromCommit->commitInfo[tid].squash) {
801 squash(tid);
802
803 if (dispatchStatus[tid] == Blocked ||
804 dispatchStatus[tid] == Unblocking) {
805 toRename->iewUnblock[tid] = true;
806 wroteToTimeBuffer = true;
807 }
808
809 dispatchStatus[tid] = Squashing;
810 fetchRedirect[tid] = false;
811 return;
812 }
813
814 if (fromCommit->commitInfo[tid].robSquashing) {
815 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
816
817 dispatchStatus[tid] = Squashing;
818 emptyRenameInsts(tid);
819 wroteToTimeBuffer = true;
820 return;
821 }
822
823 if (checkStall(tid)) {
824 block(tid);
825 dispatchStatus[tid] = Blocked;
826 return;
827 }
828
829 if (dispatchStatus[tid] == Blocked) {
830 // Status from previous cycle was blocked, but there are no more stall
831 // conditions. Switch over to unblocking.
832 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
833 tid);
834
835 dispatchStatus[tid] = Unblocking;
836
837 unblock(tid);
838
839 return;
840 }
841
842 if (dispatchStatus[tid] == Squashing) {
843 // Switch status to running if rename isn't being told to block or
844 // squash this cycle.
845 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
846 tid);
847
848 dispatchStatus[tid] = Running;
849
850 return;
851 }
852}
853
854template <class Impl>
855void
856DefaultIEW<Impl>::sortInsts()
857{
858 int insts_from_rename = fromRename->size;
859#ifdef DEBUG
860 for (ThreadID tid = 0; tid < numThreads; tid++)
861 assert(insts[tid].empty());
862#endif
863 for (int i = 0; i < insts_from_rename; ++i) {
864 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
865 }
866}
867
868template <class Impl>
869void
870DefaultIEW<Impl>::emptyRenameInsts(ThreadID tid)
871{
872 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions\n", tid);
873
874 while (!insts[tid].empty()) {
875
876 if (insts[tid].front()->isLoad() ||
877 insts[tid].front()->isStore() ) {
878 toRename->iewInfo[tid].dispatchedToLSQ++;
879 }
880
881 toRename->iewInfo[tid].dispatched++;
882
883 insts[tid].pop();
884 }
885}
886
887template <class Impl>
888void
889DefaultIEW<Impl>::wakeCPU()
890{
891 cpu->wakeCPU();
892}
893
894template <class Impl>
895void
896DefaultIEW<Impl>::activityThisCycle()
897{
898 DPRINTF(Activity, "Activity this cycle.\n");
899 cpu->activityThisCycle();
900}
901
902template <class Impl>
903inline void
904DefaultIEW<Impl>::activateStage()
905{
906 DPRINTF(Activity, "Activating stage.\n");
907 cpu->activateStage(O3CPU::IEWIdx);
908}
909
910template <class Impl>
911inline void
912DefaultIEW<Impl>::deactivateStage()
913{
914 DPRINTF(Activity, "Deactivating stage.\n");
915 cpu->deactivateStage(O3CPU::IEWIdx);
916}
917
918template<class Impl>
919void
920DefaultIEW<Impl>::dispatch(ThreadID tid)
921{
922 // If status is Running or idle,
923 // call dispatchInsts()
924 // If status is Unblocking,
925 // buffer any instructions coming from rename
926 // continue trying to empty skid buffer
927 // check if stall conditions have passed
928
929 if (dispatchStatus[tid] == Blocked) {
930 ++iewBlockCycles;
931
932 } else if (dispatchStatus[tid] == Squashing) {
933 ++iewSquashCycles;
934 }
935
936 // Dispatch should try to dispatch as many instructions as its bandwidth
937 // will allow, as long as it is not currently blocked.
938 if (dispatchStatus[tid] == Running ||
939 dispatchStatus[tid] == Idle) {
940 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
941 "dispatch.\n", tid);
942
943 dispatchInsts(tid);
944 } else if (dispatchStatus[tid] == Unblocking) {
945 // Make sure that the skid buffer has something in it if the
946 // status is unblocking.
947 assert(!skidsEmpty());
948
949 // If the status was unblocking, then instructions from the skid
950 // buffer were used. Remove those instructions and handle
951 // the rest of unblocking.
952 dispatchInsts(tid);
953
954 ++iewUnblockCycles;
955
956 if (validInstsFromRename()) {
957 // Add the current inputs to the skid buffer so they can be
958 // reprocessed when this stage unblocks.
959 skidInsert(tid);
960 }
961
962 unblock(tid);
963 }
964}
965
966template <class Impl>
967void
968DefaultIEW<Impl>::dispatchInsts(ThreadID tid)
969{
970 // Obtain instructions from skid buffer if unblocking, or queue from rename
971 // otherwise.
972 std::queue<DynInstPtr> &insts_to_dispatch =
973 dispatchStatus[tid] == Unblocking ?
974 skidBuffer[tid] : insts[tid];
975
976 int insts_to_add = insts_to_dispatch.size();
977
978 DynInstPtr inst;
979 bool add_to_iq = false;
980 int dis_num_inst = 0;
981
982 // Loop through the instructions, putting them in the instruction
983 // queue.
984 for ( ; dis_num_inst < insts_to_add &&
985 dis_num_inst < dispatchWidth;
986 ++dis_num_inst)
987 {
988 inst = insts_to_dispatch.front();
989
990 if (dispatchStatus[tid] == Unblocking) {
991 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
992 "buffer\n", tid);
993 }
994
995 // Make sure there's a valid instruction there.
996 assert(inst);
997
998 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %s [sn:%lli] [tid:%i] to "
999 "IQ.\n",
1000 tid, inst->pcState(), inst->seqNum, inst->threadNumber);
1001
1002 // Be sure to mark these instructions as ready so that the
1003 // commit stage can go ahead and execute them, and mark
1004 // them as issued so the IQ doesn't reprocess them.
1005
1006 // Check for squashed instructions.
1007 if (inst->isSquashed()) {
1008 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
1009 "not adding to IQ.\n", tid);
1010
1011 ++iewDispSquashedInsts;
1012
1013 insts_to_dispatch.pop();
1014
1015 //Tell Rename That An Instruction has been processed
1016 if (inst->isLoad() || inst->isStore()) {
1017 toRename->iewInfo[tid].dispatchedToLSQ++;
1018 }
1019 toRename->iewInfo[tid].dispatched++;
1020
1021 continue;
1022 }
1023
1024 // Check for full conditions.
1025 if (instQueue.isFull(tid)) {
1026 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1027
1028 // Call function to start blocking.
1029 block(tid);
1030
1031 // Set unblock to false. Special case where we are using
1032 // skidbuffer (unblocking) instructions but then we still
1033 // get full in the IQ.
1034 toRename->iewUnblock[tid] = false;
1035
1036 ++iewIQFullEvents;
1037 break;
1038 } else if (ldstQueue.isFull(tid)) {
1039 DPRINTF(IEW, "[tid:%i]: Issue: LSQ 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 ++iewLSQFullEvents;
1050 break;
1051 }
1052
1053 // Otherwise issue the instruction just fine.
1054 if (inst->isLoad()) {
1055 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1056 "encountered, adding to LSQ.\n", tid);
1057
1058 // Reserve a spot in the load store queue for this
1059 // memory access.
1060 ldstQueue.insertLoad(inst);
1061
1062 ++iewDispLoadInsts;
1063
1064 add_to_iq = true;
1065
1066 toRename->iewInfo[tid].dispatchedToLSQ++;
1067 } else if (inst->isStore()) {
1068 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1069 "encountered, adding to LSQ.\n", tid);
1070
1071 ldstQueue.insertStore(inst);
1072
1073 ++iewDispStoreInsts;
1074
1075 if (inst->isStoreConditional()) {
1076 // Store conditionals need to be set as "canCommit()"
1077 // so that commit can process them when they reach the
1078 // head of commit.
1079 // @todo: This is somewhat specific to Alpha.
1080 inst->setCanCommit();
1081 instQueue.insertNonSpec(inst);
1082 add_to_iq = false;
1083
1084 ++iewDispNonSpecInsts;
1085 } else {
1086 add_to_iq = true;
1087 }
1088
1089 toRename->iewInfo[tid].dispatchedToLSQ++;
1090 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1091 // Same as non-speculative stores.
1092 inst->setCanCommit();
1093 instQueue.insertBarrier(inst);
1094 add_to_iq = false;
1095 } else if (inst->isNop()) {
1096 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1097 "skipping.\n", tid);
1098
1099 inst->setIssued();
1100 inst->setExecuted();
1101 inst->setCanCommit();
1102
1103 instQueue.recordProducer(inst);
1104
1105 iewExecutedNop[tid]++;
1106
1107 add_to_iq = false;
1108 } else if (inst->isExecuted()) {
1109 assert(0 && "Instruction shouldn't be executed.\n");
1110 DPRINTF(IEW, "Issue: Executed branch encountered, "
1111 "skipping.\n");
1112
1113 inst->setIssued();
1114 inst->setCanCommit();
1115
1116 instQueue.recordProducer(inst);
1117
1118 add_to_iq = false;
1119 } else {
1120 add_to_iq = true;
1121 }
1122 if (inst->isNonSpeculative()) {
1123 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1124 "encountered, skipping.\n", tid);
1125
1126 // Same as non-speculative stores.
1127 inst->setCanCommit();
1128
1129 // Specifically insert it as nonspeculative.
1130 instQueue.insertNonSpec(inst);
1131
1132 ++iewDispNonSpecInsts;
1133
1134 add_to_iq = false;
1135 }
1136
1137 // If the instruction queue is not full, then add the
1138 // instruction.
1139 if (add_to_iq) {
1140 instQueue.insert(inst);
1141 }
1142
1143 insts_to_dispatch.pop();
1144
1145 toRename->iewInfo[tid].dispatched++;
1146
1147 ++iewDispatchedInsts;
1148
1149#if TRACING_ON
1150 inst->dispatchTick = curTick() - inst->fetchTick;
1151#endif
1152 }
1153
1154 if (!insts_to_dispatch.empty()) {
1155 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1156 block(tid);
1157 toRename->iewUnblock[tid] = false;
1158 }
1159
1160 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1161 dispatchStatus[tid] = Running;
1162
1163 updatedQueues = true;
1164 }
1165
1166 dis_num_inst = 0;
1167}
1168
1169template <class Impl>
1170void
1171DefaultIEW<Impl>::printAvailableInsts()
1172{
1173 int inst = 0;
1174
1175 std::cout << "Available Instructions: ";
1176
1177 while (fromIssue->insts[inst]) {
1178
1179 if (inst%3==0) std::cout << "\n\t";
1180
1181 std::cout << "PC: " << fromIssue->insts[inst]->pcState()
1182 << " TN: " << fromIssue->insts[inst]->threadNumber
1183 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1184
1185 inst++;
1186
1187 }
1188
1189 std::cout << "\n";
1190}
1191
1192template <class Impl>
1193void
1194DefaultIEW<Impl>::executeInsts()
1195{
1196 wbNumInst = 0;
1197 wbCycle = 0;
1198
1199 list<ThreadID>::iterator threads = activeThreads->begin();
1200 list<ThreadID>::iterator end = activeThreads->end();
1201
1202 while (threads != end) {
1203 ThreadID tid = *threads++;
1204 fetchRedirect[tid] = false;
1205 }
1206
1207 // Uncomment this if you want to see all available instructions.
1208 // @todo This doesn't actually work anymore, we should fix it.
1209// printAvailableInsts();
1210
1211 // Execute/writeback any instructions that are available.
1212 int insts_to_execute = fromIssue->size;
1213 int inst_num = 0;
1214 for (; inst_num < insts_to_execute;
1215 ++inst_num) {
1216
1217 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1218
1219 DynInstPtr inst = instQueue.getInstToExecute();
1220
1221 DPRINTF(IEW, "Execute: Processing PC %s, [tid:%i] [sn:%i].\n",
1222 inst->pcState(), inst->threadNumber,inst->seqNum);
1223
1224 // Check if the instruction is squashed; if so then skip it
1225 if (inst->isSquashed()) {
1226 DPRINTF(IEW, "Execute: Instruction was squashed. PC: %s, [tid:%i]"
1227 " [sn:%i]\n", inst->pcState(), inst->threadNumber,
1228 inst->seqNum);
1229
1230 // Consider this instruction executed so that commit can go
1231 // ahead and retire the instruction.
1232 inst->setExecuted();
1233
1234 // Not sure if I should set this here or just let commit try to
1235 // commit any squashed instructions. I like the latter a bit more.
1236 inst->setCanCommit();
1237
1238 ++iewExecSquashedInsts;
1239
1240 decrWb(inst->seqNum);
1241 continue;
1242 }
1243
1244 Fault fault = NoFault;
1245
1246 // Execute instruction.
1247 // Note that if the instruction faults, it will be handled
1248 // at the commit stage.
1249 if (inst->isMemRef()) {
1250 DPRINTF(IEW, "Execute: Calculating address for memory "
1251 "reference.\n");
1252
1253 // Tell the LDSTQ to execute this instruction (if it is a load).
1254 if (inst->isLoad()) {
1255 // Loads will mark themselves as executed, and their writeback
1256 // event adds the instruction to the queue to commit
1257 fault = ldstQueue.executeLoad(inst);
1258
1259 if (inst->isTranslationDelayed() &&
1260 fault == NoFault) {
1261 // A hw page table walk is currently going on; the
1262 // instruction must be deferred.
1263 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1264 "load.\n");
1265 instQueue.deferMemInst(inst);
1266 continue;
1267 }
1268
1269 if (inst->isDataPrefetch() || inst->isInstPrefetch()) {
1270 inst->fault = NoFault;
1271 }
1272 } else if (inst->isStore()) {
1273 fault = ldstQueue.executeStore(inst);
1274
1275 if (inst->isTranslationDelayed() &&
1276 fault == NoFault) {
1277 // A hw page table walk is currently going on; the
1278 // instruction must be deferred.
1279 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1280 "store.\n");
1281 instQueue.deferMemInst(inst);
1282 continue;
1283 }
1284
1285 // If the store had a fault then it may not have a mem req
1286 if (fault != NoFault || inst->readPredicate() == false ||
1287 !inst->isStoreConditional()) {
1288 // If the instruction faulted, then we need to send it along
1289 // to commit without the instruction completing.
1290 // Send this instruction to commit, also make sure iew stage
1291 // realizes there is activity.
1292 inst->setExecuted();
1293 instToCommit(inst);
1294 activityThisCycle();
1295 }
1296
1297 // Store conditionals will mark themselves as
1298 // executed, and their writeback event will add the
1299 // instruction to the queue to commit.
1300 } else {
1301 panic("Unexpected memory type!\n");
1302 }
1303
1304 } else {
1305 // If the instruction has already faulted, then skip executing it.
1306 // Such case can happen when it faulted during ITLB translation.
1307 // If we execute the instruction (even if it's a nop) the fault
1308 // will be replaced and we will lose it.
1309 if (inst->getFault() == NoFault) {
1310 inst->execute();
1311 if (inst->readPredicate() == false)
1312 inst->forwardOldRegs();
1313 }
1314
1315 inst->setExecuted();
1316
1317 instToCommit(inst);
1318 }
1319
1320 updateExeInstStats(inst);
1321
1322 // Check if branch prediction was correct, if not then we need
1323 // to tell commit to squash in flight instructions. Only
1324 // handle this if there hasn't already been something that
1325 // redirects fetch in this group of instructions.
1326
1327 // This probably needs to prioritize the redirects if a different
1328 // scheduler is used. Currently the scheduler schedules the oldest
1329 // instruction first, so the branch resolution order will be correct.
1330 ThreadID tid = inst->threadNumber;
1331
1332 if (!fetchRedirect[tid] ||
1333 !toCommit->squash[tid] ||
1334 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1335
1336 // Prevent testing for misprediction on load instructions,
1337 // that have not been executed.
1338 bool loadNotExecuted = !inst->isExecuted() && inst->isLoad();
1339
1340 if (inst->mispredicted() && !loadNotExecuted) {
1341 fetchRedirect[tid] = true;
1342
1343 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1344 DPRINTF(IEW, "Predicted target was PC: %s.\n",
1345 inst->readPredTarg());
1346 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %s.\n",
1347 inst->pcState());
1348 // If incorrect, then signal the ROB that it must be squashed.
1349 squashDueToBranch(inst, tid);
1350
1351 if (inst->readPredTaken()) {
1352 predictedTakenIncorrect++;
1353 } else {
1354 predictedNotTakenIncorrect++;
1355 }
1356 } else if (ldstQueue.violation(tid)) {
1357 assert(inst->isMemRef());
1358 // If there was an ordering violation, then get the
1359 // DynInst that caused the violation. Note that this
1360 // clears the violation signal.
1361 DynInstPtr violator;
1362 violator = ldstQueue.getMemDepViolator(tid);
1363
1364 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: %s "
1365 "[sn:%lli], inst PC: %s [sn:%lli]. Addr is: %#x.\n",
1366 violator->pcState(), violator->seqNum,
1367 inst->pcState(), inst->seqNum, inst->physEffAddr);
1368
1369 fetchRedirect[tid] = true;
1370
1371 // Tell the instruction queue that a violation has occured.
1372 instQueue.violation(inst, violator);
1373
1374 // Squash.
1375 squashDueToMemOrder(violator, tid);
1376
1377 ++memOrderViolationEvents;
1378 } else if (ldstQueue.loadBlocked(tid) &&
1379 !ldstQueue.isLoadBlockedHandled(tid)) {
1380 fetchRedirect[tid] = true;
1381
1382 DPRINTF(IEW, "Load operation couldn't execute because the "
1383 "memory system is blocked. PC: %s [sn:%lli]\n",
1384 inst->pcState(), inst->seqNum);
1385
1386 squashDueToMemBlocked(inst, tid);
1387 }
1388 } else {
1389 // Reset any state associated with redirects that will not
1390 // be used.
1391 if (ldstQueue.violation(tid)) {
1392 assert(inst->isMemRef());
1393
1394 DynInstPtr violator = ldstQueue.getMemDepViolator(tid);
1395
1396 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1397 "%s, inst PC: %s. Addr is: %#x.\n",
1398 violator->pcState(), inst->pcState(),
1399 inst->physEffAddr);
1400 DPRINTF(IEW, "Violation will not be handled because "
1401 "already squashing\n");
1402
1403 ++memOrderViolationEvents;
1404 }
1405 if (ldstQueue.loadBlocked(tid) &&
1406 !ldstQueue.isLoadBlockedHandled(tid)) {
1407 DPRINTF(IEW, "Load operation couldn't execute because the "
1408 "memory system is blocked. PC: %s [sn:%lli]\n",
1409 inst->pcState(), inst->seqNum);
1410 DPRINTF(IEW, "Blocked load will not be handled because "
1411 "already squashing\n");
1412
1413 ldstQueue.setLoadBlockedHandled(tid);
1414 }
1415
1416 }
1417 }
1418
1419 // Update and record activity if we processed any instructions.
1420 if (inst_num) {
1421 if (exeStatus == Idle) {
1422 exeStatus = Running;
1423 }
1424
1425 updatedQueues = true;
1426
1427 cpu->activityThisCycle();
1428 }
1429
1430 // Need to reset this in case a writeback event needs to write into the
1431 // iew queue. That way the writeback event will write into the correct
1432 // spot in the queue.
1433 wbNumInst = 0;
1434
1435}
1436
1437template <class Impl>
1438void
1439DefaultIEW<Impl>::writebackInsts()
1440{
1441 // Loop through the head of the time buffer and wake any
1442 // dependents. These instructions are about to write back. Also
1443 // mark scoreboard that this instruction is finally complete.
1444 // Either have IEW have direct access to scoreboard, or have this
1445 // as part of backwards communication.
1446 for (int inst_num = 0; inst_num < wbWidth &&
1447 toCommit->insts[inst_num]; inst_num++) {
1448 DynInstPtr inst = toCommit->insts[inst_num];
1449 ThreadID tid = inst->threadNumber;
1450
1451 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %s.\n",
1452 inst->seqNum, inst->pcState());
1453
1454 iewInstsToCommit[tid]++;
1455
1456 // Some instructions will be sent to commit without having
1457 // executed because they need commit to handle them.
1458 // E.g. Uncached loads have not actually executed when they
1459 // are first sent to commit. Instead commit must tell the LSQ
1460 // when it's ready to execute the uncached load.
1461 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1462 int dependents = instQueue.wakeDependents(inst);
1463
1464 for (int i = 0; i < inst->numDestRegs(); i++) {
1465 //mark as Ready
1466 DPRINTF(IEW,"Setting Destination Register %i\n",
1467 inst->renamedDestRegIdx(i));
1468 scoreboard->setReg(inst->renamedDestRegIdx(i));
1469 }
1470
1471 if (dependents) {
1472 producerInst[tid]++;
1473 consumerInst[tid]+= dependents;
1474 }
1475 writebackCount[tid]++;
1476 }
1477
1478 decrWb(inst->seqNum);
1479 }
1480}
1481
1482template<class Impl>
1483void
1484DefaultIEW<Impl>::tick()
1485{
1486 wbNumInst = 0;
1487 wbCycle = 0;
1488
1489 wroteToTimeBuffer = false;
1490 updatedQueues = false;
1491
1492 sortInsts();
1493
1494 // Free function units marked as being freed this cycle.
1495 fuPool->processFreeUnits();
1496
1497 list<ThreadID>::iterator threads = activeThreads->begin();
1498 list<ThreadID>::iterator end = activeThreads->end();
1499
1500 // Check stall and squash signals, dispatch any instructions.
1501 while (threads != end) {
1502 ThreadID tid = *threads++;
1503
1504 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1505
1506 checkSignalsAndUpdate(tid);
1507 dispatch(tid);
1508 }
1509
1510 if (exeStatus != Squashing) {
1511 executeInsts();
1512
1513 writebackInsts();
1514
1515 // Have the instruction queue try to schedule any ready instructions.
1516 // (In actuality, this scheduling is for instructions that will
1517 // be executed next cycle.)
1518 instQueue.scheduleReadyInsts();
1519
1520 // Also should advance its own time buffers if the stage ran.
1521 // Not the best place for it, but this works (hopefully).
1522 issueToExecQueue.advance();
1523 }
1524
1525 bool broadcast_free_entries = false;
1526
1527 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1528 exeStatus = Idle;
1529 updateLSQNextCycle = false;
1530
1531 broadcast_free_entries = true;
1532 }
1533
1534 // Writeback any stores using any leftover bandwidth.
1535 ldstQueue.writebackStores();
1536
1537 // Check the committed load/store signals to see if there's a load
1538 // or store to commit. Also check if it's being told to execute a
1539 // nonspeculative instruction.
1540 // This is pretty inefficient...
1541
1542 threads = activeThreads->begin();
1543 while (threads != end) {
1544 ThreadID tid = (*threads++);
1545
1546 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1547
1548 // Update structures based on instructions committed.
1549 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1550 !fromCommit->commitInfo[tid].squash &&
1551 !fromCommit->commitInfo[tid].robSquashing) {
1552
1553 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1554
1555 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1556
1557 updateLSQNextCycle = true;
1558 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1559 }
1560
1561 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1562
1563 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1564 if (fromCommit->commitInfo[tid].uncached) {
1565 instQueue.replayMemInst(fromCommit->commitInfo[tid].uncachedLoad);
1566 fromCommit->commitInfo[tid].uncachedLoad->setAtCommit();
1567 } else {
1568 instQueue.scheduleNonSpec(
1569 fromCommit->commitInfo[tid].nonSpecSeqNum);
1570 }
1571 }
1572
1573 if (broadcast_free_entries) {
1574 toFetch->iewInfo[tid].iqCount =
1575 instQueue.getCount(tid);
1576 toFetch->iewInfo[tid].ldstqCount =
1577 ldstQueue.getCount(tid);
1578
1579 toRename->iewInfo[tid].usedIQ = true;
1580 toRename->iewInfo[tid].freeIQEntries =
1581 instQueue.numFreeEntries();
1582 toRename->iewInfo[tid].usedLSQ = true;
1583 toRename->iewInfo[tid].freeLSQEntries =
1584 ldstQueue.numFreeEntries(tid);
1585
1586 wroteToTimeBuffer = true;
1587 }
1588
1589 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1590 tid, toRename->iewInfo[tid].dispatched);
1591 }
1592
1593 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1594 "LSQ has %i free entries.\n",
1595 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1596 ldstQueue.numFreeEntries());
1597
1598 updateStatus();
1599
1600 if (wroteToTimeBuffer) {
1601 DPRINTF(Activity, "Activity this cycle.\n");
1602 cpu->activityThisCycle();
1603 }
1604}
1605
1606template <class Impl>
1607void
1608DefaultIEW<Impl>::updateExeInstStats(DynInstPtr &inst)
1609{
1610 ThreadID tid = inst->threadNumber;
1611
1612 iewExecutedInsts++;
1613
1614#if TRACING_ON
1614 inst->completeTick = curTick() - inst->fetchTick;
1615 if (DTRACE(O3PipeView)) {
1616 inst->completeTick = curTick() - inst->fetchTick;
1617 }
1615#endif
1616
1617 //
1618 // Control operations
1619 //
1620 if (inst->isControl())
1621 iewExecutedBranches[tid]++;
1622
1623 //
1624 // Memory operations
1625 //
1626 if (inst->isMemRef()) {
1627 iewExecutedRefs[tid]++;
1628
1629 if (inst->isLoad()) {
1630 iewExecLoadInsts[tid]++;
1631 }
1632 }
1633}
1634
1635template <class Impl>
1636void
1637DefaultIEW<Impl>::checkMisprediction(DynInstPtr &inst)
1638{
1639 ThreadID tid = inst->threadNumber;
1640
1641 if (!fetchRedirect[tid] ||
1642 !toCommit->squash[tid] ||
1643 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1644
1645 if (inst->mispredicted()) {
1646 fetchRedirect[tid] = true;
1647
1648 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1649 DPRINTF(IEW, "Predicted target was PC:%#x, NPC:%#x.\n",
1650 inst->predInstAddr(), inst->predNextInstAddr());
1651 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x,"
1652 " NPC: %#x.\n", inst->nextInstAddr(),
1653 inst->nextInstAddr());
1654 // If incorrect, then signal the ROB that it must be squashed.
1655 squashDueToBranch(inst, tid);
1656
1657 if (inst->readPredTaken()) {
1658 predictedTakenIncorrect++;
1659 } else {
1660 predictedNotTakenIncorrect++;
1661 }
1662 }
1663 }
1664}
1618#endif
1619
1620 //
1621 // Control operations
1622 //
1623 if (inst->isControl())
1624 iewExecutedBranches[tid]++;
1625
1626 //
1627 // Memory operations
1628 //
1629 if (inst->isMemRef()) {
1630 iewExecutedRefs[tid]++;
1631
1632 if (inst->isLoad()) {
1633 iewExecLoadInsts[tid]++;
1634 }
1635 }
1636}
1637
1638template <class Impl>
1639void
1640DefaultIEW<Impl>::checkMisprediction(DynInstPtr &inst)
1641{
1642 ThreadID tid = inst->threadNumber;
1643
1644 if (!fetchRedirect[tid] ||
1645 !toCommit->squash[tid] ||
1646 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1647
1648 if (inst->mispredicted()) {
1649 fetchRedirect[tid] = true;
1650
1651 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1652 DPRINTF(IEW, "Predicted target was PC:%#x, NPC:%#x.\n",
1653 inst->predInstAddr(), inst->predNextInstAddr());
1654 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x,"
1655 " NPC: %#x.\n", inst->nextInstAddr(),
1656 inst->nextInstAddr());
1657 // If incorrect, then signal the ROB that it must be squashed.
1658 squashDueToBranch(inst, tid);
1659
1660 if (inst->readPredTaken()) {
1661 predictedTakenIncorrect++;
1662 } else {
1663 predictedNotTakenIncorrect++;
1664 }
1665 }
1666 }
1667}