iew_impl.hh (3125:febd811bccc6) iew_impl.hh (3126:756092c6383c)
1/*
2 * Copyright (c) 2004-2006 The Regents of The University of Michigan
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are
7 * met: redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer;
9 * redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution;
12 * neither the name of the copyright holders nor the names of its
13 * contributors may be used to endorse or promote products derived from
14 * this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Authors: Kevin Lim
29 */
30
31// @todo: Fix the instantaneous communication among all the stages within
32// iew. There's a clear delay between issue and execute, yet backwards
33// communication happens simultaneously.
34
35#include <queue>
36
37#include "base/timebuf.hh"
38#include "cpu/o3/fu_pool.hh"
39#include "cpu/o3/iew.hh"
40
41template<class Impl>
42DefaultIEW<Impl>::DefaultIEW(Params *params)
43 : issueToExecQueue(params->backComSize, params->forwardComSize),
44 instQueue(params),
45 ldstQueue(params),
46 fuPool(params->fuPool),
47 commitToIEWDelay(params->commitToIEWDelay),
48 renameToIEWDelay(params->renameToIEWDelay),
49 issueToExecuteDelay(params->issueToExecuteDelay),
50 dispatchWidth(params->dispatchWidth),
51 issueWidth(params->issueWidth),
52 wbOutstanding(0),
53 wbWidth(params->wbWidth),
54 numThreads(params->numberOfThreads),
55 switchedOut(false)
56{
57 _status = Active;
58 exeStatus = Running;
59 wbStatus = Idle;
60
61 // Setup wire to read instructions coming from issue.
62 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
63
64 // Instruction queue needs the queue between issue and execute.
65 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
66
67 instQueue.setIEW(this);
68 ldstQueue.setIEW(this);
69
70 for (int i=0; i < numThreads; i++) {
71 dispatchStatus[i] = Running;
72 stalls[i].commit = false;
73 fetchRedirect[i] = false;
74 bdelayDoneSeqNum[i] = 0;
75 }
76
77 wbMax = wbWidth * params->wbDepth;
78
79 updateLSQNextCycle = false;
80
81 ableToIssue = true;
82
83 skidBufferMax = (3 * (renameToIEWDelay * params->renameWidth)) + issueWidth;
84}
85
86template <class Impl>
87std::string
88DefaultIEW<Impl>::name() const
89{
90 return cpu->name() + ".iew";
91}
92
93template <class Impl>
94void
95DefaultIEW<Impl>::regStats()
96{
97 using namespace Stats;
98
99 instQueue.regStats();
100 ldstQueue.regStats();
101
102 iewIdleCycles
103 .name(name() + ".iewIdleCycles")
104 .desc("Number of cycles IEW is idle");
105
106 iewSquashCycles
107 .name(name() + ".iewSquashCycles")
108 .desc("Number of cycles IEW is squashing");
109
110 iewBlockCycles
111 .name(name() + ".iewBlockCycles")
112 .desc("Number of cycles IEW is blocking");
113
114 iewUnblockCycles
115 .name(name() + ".iewUnblockCycles")
116 .desc("Number of cycles IEW is unblocking");
117
118 iewDispatchedInsts
119 .name(name() + ".iewDispatchedInsts")
120 .desc("Number of instructions dispatched to IQ");
121
122 iewDispSquashedInsts
123 .name(name() + ".iewDispSquashedInsts")
124 .desc("Number of squashed instructions skipped by dispatch");
125
126 iewDispLoadInsts
127 .name(name() + ".iewDispLoadInsts")
128 .desc("Number of dispatched load instructions");
129
130 iewDispStoreInsts
131 .name(name() + ".iewDispStoreInsts")
132 .desc("Number of dispatched store instructions");
133
134 iewDispNonSpecInsts
135 .name(name() + ".iewDispNonSpecInsts")
136 .desc("Number of dispatched non-speculative instructions");
137
138 iewIQFullEvents
139 .name(name() + ".iewIQFullEvents")
140 .desc("Number of times the IQ has become full, causing a stall");
141
142 iewLSQFullEvents
143 .name(name() + ".iewLSQFullEvents")
144 .desc("Number of times the LSQ has become full, causing a stall");
145
146 memOrderViolationEvents
147 .name(name() + ".memOrderViolationEvents")
148 .desc("Number of memory order violations");
149
150 predictedTakenIncorrect
151 .name(name() + ".predictedTakenIncorrect")
152 .desc("Number of branches that were predicted taken incorrectly");
153
154 predictedNotTakenIncorrect
155 .name(name() + ".predictedNotTakenIncorrect")
156 .desc("Number of branches that were predicted not taken incorrectly");
157
158 branchMispredicts
159 .name(name() + ".branchMispredicts")
160 .desc("Number of branch mispredicts detected at execute");
161
162 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
163
164 iewExecutedInsts
165 .name(name() + ".iewExecutedInsts")
166 .desc("Number of executed instructions");
167
168 iewExecLoadInsts
169 .init(cpu->number_of_threads)
170 .name(name() + ".iewExecLoadInsts")
171 .desc("Number of load instructions executed")
172 .flags(total);
173
174 iewExecSquashedInsts
175 .name(name() + ".iewExecSquashedInsts")
176 .desc("Number of squashed instructions skipped in execute");
177
178 iewExecutedSwp
179 .init(cpu->number_of_threads)
180 .name(name() + ".EXEC:swp")
181 .desc("number of swp insts executed")
182 .flags(total);
183
184 iewExecutedNop
185 .init(cpu->number_of_threads)
186 .name(name() + ".EXEC:nop")
187 .desc("number of nop insts executed")
188 .flags(total);
189
190 iewExecutedRefs
191 .init(cpu->number_of_threads)
192 .name(name() + ".EXEC:refs")
193 .desc("number of memory reference insts executed")
194 .flags(total);
195
196 iewExecutedBranches
197 .init(cpu->number_of_threads)
198 .name(name() + ".EXEC:branches")
199 .desc("Number of branches executed")
200 .flags(total);
201
202 iewExecStoreInsts
203 .name(name() + ".EXEC:stores")
204 .desc("Number of stores executed")
205 .flags(total);
206 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
207
208 iewExecRate
209 .name(name() + ".EXEC:rate")
210 .desc("Inst execution rate")
211 .flags(total);
212
213 iewExecRate = iewExecutedInsts / cpu->numCycles;
214
215 iewInstsToCommit
216 .init(cpu->number_of_threads)
217 .name(name() + ".WB:sent")
218 .desc("cumulative count of insts sent to commit")
219 .flags(total);
220
221 writebackCount
222 .init(cpu->number_of_threads)
223 .name(name() + ".WB:count")
224 .desc("cumulative count of insts written-back")
225 .flags(total);
226
227 producerInst
228 .init(cpu->number_of_threads)
229 .name(name() + ".WB:producers")
230 .desc("num instructions producing a value")
231 .flags(total);
232
233 consumerInst
234 .init(cpu->number_of_threads)
235 .name(name() + ".WB:consumers")
236 .desc("num instructions consuming a value")
237 .flags(total);
238
239 wbPenalized
240 .init(cpu->number_of_threads)
241 .name(name() + ".WB:penalized")
242 .desc("number of instrctions required to write to 'other' IQ")
243 .flags(total);
244
245 wbPenalizedRate
246 .name(name() + ".WB:penalized_rate")
247 .desc ("fraction of instructions written-back that wrote to 'other' IQ")
248 .flags(total);
249
250 wbPenalizedRate = wbPenalized / writebackCount;
251
252 wbFanout
253 .name(name() + ".WB:fanout")
254 .desc("average fanout of values written-back")
255 .flags(total);
256
257 wbFanout = producerInst / consumerInst;
258
259 wbRate
260 .name(name() + ".WB:rate")
261 .desc("insts written-back per cycle")
262 .flags(total);
263 wbRate = writebackCount / cpu->numCycles;
264}
265
266template<class Impl>
267void
268DefaultIEW<Impl>::initStage()
269{
270 for (int tid=0; tid < numThreads; tid++) {
271 toRename->iewInfo[tid].usedIQ = true;
272 toRename->iewInfo[tid].freeIQEntries =
273 instQueue.numFreeEntries(tid);
274
275 toRename->iewInfo[tid].usedLSQ = true;
276 toRename->iewInfo[tid].freeLSQEntries =
277 ldstQueue.numFreeEntries(tid);
278 }
279}
280
281template<class Impl>
282void
283DefaultIEW<Impl>::setCPU(O3CPU *cpu_ptr)
284{
285 DPRINTF(IEW, "Setting CPU pointer.\n");
286 cpu = cpu_ptr;
287
288 instQueue.setCPU(cpu_ptr);
289 ldstQueue.setCPU(cpu_ptr);
290
291 cpu->activateStage(O3CPU::IEWIdx);
292}
293
294template<class Impl>
295void
296DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
297{
298 DPRINTF(IEW, "Setting time buffer pointer.\n");
299 timeBuffer = tb_ptr;
300
301 // Setup wire to read information from time buffer, from commit.
302 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
303
304 // Setup wire to write information back to previous stages.
305 toRename = timeBuffer->getWire(0);
306
307 toFetch = timeBuffer->getWire(0);
308
309 // Instruction queue also needs main time buffer.
310 instQueue.setTimeBuffer(tb_ptr);
311}
312
313template<class Impl>
314void
315DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
316{
317 DPRINTF(IEW, "Setting rename queue pointer.\n");
318 renameQueue = rq_ptr;
319
320 // Setup wire to read information from rename queue.
321 fromRename = renameQueue->getWire(-renameToIEWDelay);
322}
323
324template<class Impl>
325void
326DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
327{
328 DPRINTF(IEW, "Setting IEW queue pointer.\n");
329 iewQueue = iq_ptr;
330
331 // Setup wire to write instructions to commit.
332 toCommit = iewQueue->getWire(0);
333}
334
335template<class Impl>
336void
337DefaultIEW<Impl>::setActiveThreads(std::list<unsigned> *at_ptr)
338{
339 DPRINTF(IEW, "Setting active threads list pointer.\n");
340 activeThreads = at_ptr;
341
342 ldstQueue.setActiveThreads(at_ptr);
343 instQueue.setActiveThreads(at_ptr);
344}
345
346template<class Impl>
347void
348DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
349{
350 DPRINTF(IEW, "Setting scoreboard pointer.\n");
351 scoreboard = sb_ptr;
352}
353
354template <class Impl>
355bool
356DefaultIEW<Impl>::drain()
357{
358 // IEW is ready to drain at any time.
359 cpu->signalDrained();
360 return true;
361}
362
363template <class Impl>
364void
365DefaultIEW<Impl>::resume()
366{
367}
368
369template <class Impl>
370void
371DefaultIEW<Impl>::switchOut()
372{
373 // Clear any state.
374 switchedOut = true;
375 assert(insts[0].empty());
376 assert(skidBuffer[0].empty());
377
378 instQueue.switchOut();
379 ldstQueue.switchOut();
380 fuPool->switchOut();
381
382 for (int i = 0; i < numThreads; i++) {
383 while (!insts[i].empty())
384 insts[i].pop();
385 while (!skidBuffer[i].empty())
386 skidBuffer[i].pop();
387 }
388}
389
390template <class Impl>
391void
392DefaultIEW<Impl>::takeOverFrom()
393{
394 // Reset all state.
395 _status = Active;
396 exeStatus = Running;
397 wbStatus = Idle;
398 switchedOut = false;
399
400 instQueue.takeOverFrom();
401 ldstQueue.takeOverFrom();
402 fuPool->takeOverFrom();
403
404 initStage();
405 cpu->activityThisCycle();
406
407 for (int i=0; i < numThreads; i++) {
408 dispatchStatus[i] = Running;
409 stalls[i].commit = false;
410 fetchRedirect[i] = false;
411 }
412
413 updateLSQNextCycle = false;
414
415 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
416 issueToExecQueue.advance();
417 }
418}
419
420template<class Impl>
421void
422DefaultIEW<Impl>::squash(unsigned tid)
423{
424 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n",
425 tid);
426
427 // Tell the IQ to start squashing.
428 instQueue.squash(tid);
429
430 // Tell the LDSTQ to start squashing.
431#if ISA_HAS_DELAY_SLOT
432 ldstQueue.squash(fromCommit->commitInfo[tid].bdelayDoneSeqNum, tid);
433#else
434 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
435#endif
436 updatedQueues = true;
437
438 // Clear the skid buffer in case it has any data in it.
439 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
440 tid, fromCommit->commitInfo[tid].bdelayDoneSeqNum);
441
442 while (!skidBuffer[tid].empty()) {
443#if ISA_HAS_DELAY_SLOT
444 if (skidBuffer[tid].front()->seqNum <=
445 fromCommit->commitInfo[tid].bdelayDoneSeqNum) {
446 DPRINTF(IEW, "[tid:%i]: Cannot remove skidbuffer instructions "
447 "that occur before delay slot [sn:%i].\n",
448 fromCommit->commitInfo[tid].bdelayDoneSeqNum,
449 tid);
450 break;
451 } else {
452 DPRINTF(IEW, "[tid:%i]: Removing instruction [sn:%i] from "
453 "skidBuffer.\n", tid, skidBuffer[tid].front()->seqNum);
454 }
455#endif
456 if (skidBuffer[tid].front()->isLoad() ||
457 skidBuffer[tid].front()->isStore() ) {
458 toRename->iewInfo[tid].dispatchedToLSQ++;
459 }
460
461 toRename->iewInfo[tid].dispatched++;
462
463 skidBuffer[tid].pop();
464 }
465
466 bdelayDoneSeqNum[tid] = fromCommit->commitInfo[tid].bdelayDoneSeqNum;
467
468 emptyRenameInsts(tid);
469}
470
471template<class Impl>
472void
473DefaultIEW<Impl>::squashDueToBranch(DynInstPtr &inst, unsigned tid)
474{
475 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %#x "
476 "[sn:%i].\n", tid, inst->readPC(), inst->seqNum);
477
478 toCommit->squash[tid] = true;
479 toCommit->squashedSeqNum[tid] = inst->seqNum;
480 toCommit->mispredPC[tid] = inst->readPC();
481 toCommit->branchMispredict[tid] = true;
482
483#if ISA_HAS_DELAY_SLOT
484 bool branch_taken = inst->readNextNPC() !=
485 (inst->readNextPC() + sizeof(TheISA::MachInst));
486
487 toCommit->branchTaken[tid] = branch_taken;
488
489 toCommit->condDelaySlotBranch[tid] = inst->isCondDelaySlot();
490
491 if (inst->isCondDelaySlot() && branch_taken) {
492 toCommit->nextPC[tid] = inst->readNextPC();
493 } else {
494 toCommit->nextPC[tid] = inst->readNextNPC();
495 }
496#else
497 toCommit->branchTaken[tid] = inst->readNextPC() !=
498 (inst->readPC() + sizeof(TheISA::MachInst));
499 toCommit->nextPC[tid] = inst->readNextPC();
500#endif
501
502 toCommit->includeSquashInst[tid] = false;
503
504 wroteToTimeBuffer = true;
505}
506
507template<class Impl>
508void
509DefaultIEW<Impl>::squashDueToMemOrder(DynInstPtr &inst, unsigned tid)
510{
511 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, "
512 "PC: %#x [sn:%i].\n", tid, inst->readPC(), inst->seqNum);
513
514 toCommit->squash[tid] = true;
515 toCommit->squashedSeqNum[tid] = inst->seqNum;
516 toCommit->nextPC[tid] = inst->readNextPC();
517
518 toCommit->includeSquashInst[tid] = false;
519
520 wroteToTimeBuffer = true;
521}
522
523template<class Impl>
524void
525DefaultIEW<Impl>::squashDueToMemBlocked(DynInstPtr &inst, unsigned tid)
526{
527 DPRINTF(IEW, "[tid:%i]: Memory blocked, squashing load and younger insts, "
528 "PC: %#x [sn:%i].\n", tid, inst->readPC(), inst->seqNum);
529
530 toCommit->squash[tid] = true;
531 toCommit->squashedSeqNum[tid] = inst->seqNum;
532 toCommit->nextPC[tid] = inst->readPC();
533
534 // Must include the broadcasted SN in the squash.
535 toCommit->includeSquashInst[tid] = true;
536
537 ldstQueue.setLoadBlockedHandled(tid);
538
539 wroteToTimeBuffer = true;
540}
541
542template<class Impl>
543void
544DefaultIEW<Impl>::block(unsigned tid)
545{
546 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
547
548 if (dispatchStatus[tid] != Blocked &&
549 dispatchStatus[tid] != Unblocking) {
550 toRename->iewBlock[tid] = true;
551 wroteToTimeBuffer = true;
552 }
553
554 // Add the current inputs to the skid buffer so they can be
555 // reprocessed when this stage unblocks.
556 skidInsert(tid);
557
558 dispatchStatus[tid] = Blocked;
559}
560
561template<class Impl>
562void
563DefaultIEW<Impl>::unblock(unsigned tid)
564{
565 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
566 "buffer %u.\n",tid, tid);
567
568 // If the skid bufffer is empty, signal back to previous stages to unblock.
569 // Also switch status to running.
570 if (skidBuffer[tid].empty()) {
571 toRename->iewUnblock[tid] = true;
572 wroteToTimeBuffer = true;
573 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
574 dispatchStatus[tid] = Running;
575 }
576}
577
578template<class Impl>
579void
580DefaultIEW<Impl>::wakeDependents(DynInstPtr &inst)
581{
582 instQueue.wakeDependents(inst);
583}
584
585template<class Impl>
586void
587DefaultIEW<Impl>::rescheduleMemInst(DynInstPtr &inst)
588{
589 instQueue.rescheduleMemInst(inst);
590}
591
592template<class Impl>
593void
594DefaultIEW<Impl>::replayMemInst(DynInstPtr &inst)
595{
596 instQueue.replayMemInst(inst);
597}
598
599template<class Impl>
600void
601DefaultIEW<Impl>::instToCommit(DynInstPtr &inst)
602{
603 // First check the time slot that this instruction will write
604 // to. If there are free write ports at the time, then go ahead
605 // and write the instruction to that time. If there are not,
606 // keep looking back to see where's the first time there's a
607 // free slot.
608 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
609 ++wbNumInst;
610 if (wbNumInst == wbWidth) {
611 ++wbCycle;
612 wbNumInst = 0;
613 }
614
615 assert((wbCycle * wbWidth + wbNumInst) <= wbMax);
616 }
617
618 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
619 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
620 // Add finished instruction to queue to commit.
621 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
622 (*iewQueue)[wbCycle].size++;
623}
624
625template <class Impl>
626unsigned
627DefaultIEW<Impl>::validInstsFromRename()
628{
629 unsigned inst_count = 0;
630
631 for (int i=0; i<fromRename->size; i++) {
632 if (!fromRename->insts[i]->isSquashed())
633 inst_count++;
634 }
635
636 return inst_count;
637}
638
639template<class Impl>
640void
641DefaultIEW<Impl>::skidInsert(unsigned tid)
642{
643 DynInstPtr inst = NULL;
644
645 while (!insts[tid].empty()) {
646 inst = insts[tid].front();
647
648 insts[tid].pop();
649
650 DPRINTF(Decode,"[tid:%i]: Inserting [sn:%lli] PC:%#x into "
651 "dispatch skidBuffer %i\n",tid, inst->seqNum,
652 inst->readPC(),tid);
653
654 skidBuffer[tid].push(inst);
655 }
656
657 assert(skidBuffer[tid].size() <= skidBufferMax &&
658 "Skidbuffer Exceeded Max Size");
659}
660
661template<class Impl>
662int
663DefaultIEW<Impl>::skidCount()
664{
665 int max=0;
666
667 std::list<unsigned>::iterator threads = (*activeThreads).begin();
668
669 while (threads != (*activeThreads).end()) {
670 unsigned thread_count = skidBuffer[*threads++].size();
671 if (max < thread_count)
672 max = thread_count;
673 }
674
675 return max;
676}
677
678template<class Impl>
679bool
680DefaultIEW<Impl>::skidsEmpty()
681{
682 std::list<unsigned>::iterator threads = (*activeThreads).begin();
683
684 while (threads != (*activeThreads).end()) {
685 if (!skidBuffer[*threads++].empty())
686 return false;
687 }
688
689 return true;
690}
691
692template <class Impl>
693void
694DefaultIEW<Impl>::updateStatus()
695{
696 bool any_unblocking = false;
697
698 std::list<unsigned>::iterator threads = (*activeThreads).begin();
699
700 threads = (*activeThreads).begin();
701
702 while (threads != (*activeThreads).end()) {
703 unsigned tid = *threads++;
704
705 if (dispatchStatus[tid] == Unblocking) {
706 any_unblocking = true;
707 break;
708 }
709 }
710
711 // If there are no ready instructions waiting to be scheduled by the IQ,
712 // and there's no stores waiting to write back, and dispatch is not
713 // unblocking, then there is no internal activity for the IEW stage.
714 if (_status == Active && !instQueue.hasReadyInsts() &&
715 !ldstQueue.willWB() && !any_unblocking) {
716 DPRINTF(IEW, "IEW switching to idle\n");
717
718 deactivateStage();
719
720 _status = Inactive;
721 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
722 ldstQueue.willWB() ||
723 any_unblocking)) {
724 // Otherwise there is internal activity. Set to active.
725 DPRINTF(IEW, "IEW switching to active\n");
726
727 activateStage();
728
729 _status = Active;
730 }
731}
732
733template <class Impl>
734void
735DefaultIEW<Impl>::resetEntries()
736{
737 instQueue.resetEntries();
738 ldstQueue.resetEntries();
739}
740
741template <class Impl>
742void
743DefaultIEW<Impl>::readStallSignals(unsigned tid)
744{
745 if (fromCommit->commitBlock[tid]) {
746 stalls[tid].commit = true;
747 }
748
749 if (fromCommit->commitUnblock[tid]) {
750 assert(stalls[tid].commit);
751 stalls[tid].commit = false;
752 }
753}
754
755template <class Impl>
756bool
757DefaultIEW<Impl>::checkStall(unsigned tid)
758{
759 bool ret_val(false);
760
761 if (stalls[tid].commit) {
762 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
763 ret_val = true;
764 } else if (instQueue.isFull(tid)) {
765 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
766 ret_val = true;
767 } else if (ldstQueue.isFull(tid)) {
768 DPRINTF(IEW,"[tid:%i]: Stall: LSQ is full\n",tid);
769
770 if (ldstQueue.numLoads(tid) > 0 ) {
771
772 DPRINTF(IEW,"[tid:%i]: LSQ oldest load: [sn:%i] \n",
773 tid,ldstQueue.getLoadHeadSeqNum(tid));
774 }
775
776 if (ldstQueue.numStores(tid) > 0) {
777
778 DPRINTF(IEW,"[tid:%i]: LSQ oldest store: [sn:%i] \n",
779 tid,ldstQueue.getStoreHeadSeqNum(tid));
780 }
781
782 ret_val = true;
783 } else if (ldstQueue.isStalled(tid)) {
784 DPRINTF(IEW,"[tid:%i]: Stall: LSQ stall detected.\n",tid);
785 ret_val = true;
786 }
787
788 return ret_val;
789}
790
791template <class Impl>
792void
793DefaultIEW<Impl>::checkSignalsAndUpdate(unsigned tid)
794{
795 // Check if there's a squash signal, squash if there is
796 // Check stall signals, block if there is.
797 // If status was Blocked
798 // if so then go to unblocking
799 // If status was Squashing
800 // check if squashing is not high. Switch to running this cycle.
801
802 readStallSignals(tid);
803
804 if (fromCommit->commitInfo[tid].squash) {
805 squash(tid);
806
807 if (dispatchStatus[tid] == Blocked ||
808 dispatchStatus[tid] == Unblocking) {
809 toRename->iewUnblock[tid] = true;
810 wroteToTimeBuffer = true;
811 }
812
813 dispatchStatus[tid] = Squashing;
814
815 fetchRedirect[tid] = false;
816 return;
817 }
818
819 if (fromCommit->commitInfo[tid].robSquashing) {
820 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
821
822 dispatchStatus[tid] = Squashing;
823
824 emptyRenameInsts(tid);
825 wroteToTimeBuffer = true;
826 return;
827 }
828
829 if (checkStall(tid)) {
830 block(tid);
831 dispatchStatus[tid] = Blocked;
832 return;
833 }
834
835 if (dispatchStatus[tid] == Blocked) {
836 // Status from previous cycle was blocked, but there are no more stall
837 // conditions. Switch over to unblocking.
838 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
839 tid);
840
841 dispatchStatus[tid] = Unblocking;
842
843 unblock(tid);
844
845 return;
846 }
847
848 if (dispatchStatus[tid] == Squashing) {
849 // Switch status to running if rename isn't being told to block or
850 // squash this cycle.
851 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
852 tid);
853
854 dispatchStatus[tid] = Running;
855
856 return;
857 }
858}
859
860template <class Impl>
861void
862DefaultIEW<Impl>::sortInsts()
863{
864 int insts_from_rename = fromRename->size;
865#ifdef DEBUG
866#if !ISA_HAS_DELAY_SLOT
867 for (int i = 0; i < numThreads; i++)
868 assert(insts[i].empty());
869#endif
870#endif
871 for (int i = 0; i < insts_from_rename; ++i) {
872 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
873 }
874}
875
876template <class Impl>
877void
878DefaultIEW<Impl>::emptyRenameInsts(unsigned tid)
879{
880 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions until "
881 "[sn:%i].\n", tid, bdelayDoneSeqNum[tid]);
882
883 while (!insts[tid].empty()) {
884#if ISA_HAS_DELAY_SLOT
885 if (insts[tid].front()->seqNum <= bdelayDoneSeqNum[tid]) {
886 DPRINTF(IEW, "[tid:%i]: Done removing, cannot remove instruction"
887 " that occurs at or before delay slot [sn:%i].\n",
888 tid, bdelayDoneSeqNum[tid]);
889 break;
890 } else {
891 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instruction "
892 "[sn:%i].\n", tid, insts[tid].front()->seqNum);
893 }
894#endif
895
896 if (insts[tid].front()->isLoad() ||
897 insts[tid].front()->isStore() ) {
898 toRename->iewInfo[tid].dispatchedToLSQ++;
899 }
900
901 toRename->iewInfo[tid].dispatched++;
902
903 insts[tid].pop();
904 }
905}
906
907template <class Impl>
908void
1/*
2 * Copyright (c) 2004-2006 The Regents of The University of Michigan
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are
7 * met: redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer;
9 * redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution;
12 * neither the name of the copyright holders nor the names of its
13 * contributors may be used to endorse or promote products derived from
14 * this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Authors: Kevin Lim
29 */
30
31// @todo: Fix the instantaneous communication among all the stages within
32// iew. There's a clear delay between issue and execute, yet backwards
33// communication happens simultaneously.
34
35#include <queue>
36
37#include "base/timebuf.hh"
38#include "cpu/o3/fu_pool.hh"
39#include "cpu/o3/iew.hh"
40
41template<class Impl>
42DefaultIEW<Impl>::DefaultIEW(Params *params)
43 : issueToExecQueue(params->backComSize, params->forwardComSize),
44 instQueue(params),
45 ldstQueue(params),
46 fuPool(params->fuPool),
47 commitToIEWDelay(params->commitToIEWDelay),
48 renameToIEWDelay(params->renameToIEWDelay),
49 issueToExecuteDelay(params->issueToExecuteDelay),
50 dispatchWidth(params->dispatchWidth),
51 issueWidth(params->issueWidth),
52 wbOutstanding(0),
53 wbWidth(params->wbWidth),
54 numThreads(params->numberOfThreads),
55 switchedOut(false)
56{
57 _status = Active;
58 exeStatus = Running;
59 wbStatus = Idle;
60
61 // Setup wire to read instructions coming from issue.
62 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
63
64 // Instruction queue needs the queue between issue and execute.
65 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
66
67 instQueue.setIEW(this);
68 ldstQueue.setIEW(this);
69
70 for (int i=0; i < numThreads; i++) {
71 dispatchStatus[i] = Running;
72 stalls[i].commit = false;
73 fetchRedirect[i] = false;
74 bdelayDoneSeqNum[i] = 0;
75 }
76
77 wbMax = wbWidth * params->wbDepth;
78
79 updateLSQNextCycle = false;
80
81 ableToIssue = true;
82
83 skidBufferMax = (3 * (renameToIEWDelay * params->renameWidth)) + issueWidth;
84}
85
86template <class Impl>
87std::string
88DefaultIEW<Impl>::name() const
89{
90 return cpu->name() + ".iew";
91}
92
93template <class Impl>
94void
95DefaultIEW<Impl>::regStats()
96{
97 using namespace Stats;
98
99 instQueue.regStats();
100 ldstQueue.regStats();
101
102 iewIdleCycles
103 .name(name() + ".iewIdleCycles")
104 .desc("Number of cycles IEW is idle");
105
106 iewSquashCycles
107 .name(name() + ".iewSquashCycles")
108 .desc("Number of cycles IEW is squashing");
109
110 iewBlockCycles
111 .name(name() + ".iewBlockCycles")
112 .desc("Number of cycles IEW is blocking");
113
114 iewUnblockCycles
115 .name(name() + ".iewUnblockCycles")
116 .desc("Number of cycles IEW is unblocking");
117
118 iewDispatchedInsts
119 .name(name() + ".iewDispatchedInsts")
120 .desc("Number of instructions dispatched to IQ");
121
122 iewDispSquashedInsts
123 .name(name() + ".iewDispSquashedInsts")
124 .desc("Number of squashed instructions skipped by dispatch");
125
126 iewDispLoadInsts
127 .name(name() + ".iewDispLoadInsts")
128 .desc("Number of dispatched load instructions");
129
130 iewDispStoreInsts
131 .name(name() + ".iewDispStoreInsts")
132 .desc("Number of dispatched store instructions");
133
134 iewDispNonSpecInsts
135 .name(name() + ".iewDispNonSpecInsts")
136 .desc("Number of dispatched non-speculative instructions");
137
138 iewIQFullEvents
139 .name(name() + ".iewIQFullEvents")
140 .desc("Number of times the IQ has become full, causing a stall");
141
142 iewLSQFullEvents
143 .name(name() + ".iewLSQFullEvents")
144 .desc("Number of times the LSQ has become full, causing a stall");
145
146 memOrderViolationEvents
147 .name(name() + ".memOrderViolationEvents")
148 .desc("Number of memory order violations");
149
150 predictedTakenIncorrect
151 .name(name() + ".predictedTakenIncorrect")
152 .desc("Number of branches that were predicted taken incorrectly");
153
154 predictedNotTakenIncorrect
155 .name(name() + ".predictedNotTakenIncorrect")
156 .desc("Number of branches that were predicted not taken incorrectly");
157
158 branchMispredicts
159 .name(name() + ".branchMispredicts")
160 .desc("Number of branch mispredicts detected at execute");
161
162 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
163
164 iewExecutedInsts
165 .name(name() + ".iewExecutedInsts")
166 .desc("Number of executed instructions");
167
168 iewExecLoadInsts
169 .init(cpu->number_of_threads)
170 .name(name() + ".iewExecLoadInsts")
171 .desc("Number of load instructions executed")
172 .flags(total);
173
174 iewExecSquashedInsts
175 .name(name() + ".iewExecSquashedInsts")
176 .desc("Number of squashed instructions skipped in execute");
177
178 iewExecutedSwp
179 .init(cpu->number_of_threads)
180 .name(name() + ".EXEC:swp")
181 .desc("number of swp insts executed")
182 .flags(total);
183
184 iewExecutedNop
185 .init(cpu->number_of_threads)
186 .name(name() + ".EXEC:nop")
187 .desc("number of nop insts executed")
188 .flags(total);
189
190 iewExecutedRefs
191 .init(cpu->number_of_threads)
192 .name(name() + ".EXEC:refs")
193 .desc("number of memory reference insts executed")
194 .flags(total);
195
196 iewExecutedBranches
197 .init(cpu->number_of_threads)
198 .name(name() + ".EXEC:branches")
199 .desc("Number of branches executed")
200 .flags(total);
201
202 iewExecStoreInsts
203 .name(name() + ".EXEC:stores")
204 .desc("Number of stores executed")
205 .flags(total);
206 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
207
208 iewExecRate
209 .name(name() + ".EXEC:rate")
210 .desc("Inst execution rate")
211 .flags(total);
212
213 iewExecRate = iewExecutedInsts / cpu->numCycles;
214
215 iewInstsToCommit
216 .init(cpu->number_of_threads)
217 .name(name() + ".WB:sent")
218 .desc("cumulative count of insts sent to commit")
219 .flags(total);
220
221 writebackCount
222 .init(cpu->number_of_threads)
223 .name(name() + ".WB:count")
224 .desc("cumulative count of insts written-back")
225 .flags(total);
226
227 producerInst
228 .init(cpu->number_of_threads)
229 .name(name() + ".WB:producers")
230 .desc("num instructions producing a value")
231 .flags(total);
232
233 consumerInst
234 .init(cpu->number_of_threads)
235 .name(name() + ".WB:consumers")
236 .desc("num instructions consuming a value")
237 .flags(total);
238
239 wbPenalized
240 .init(cpu->number_of_threads)
241 .name(name() + ".WB:penalized")
242 .desc("number of instrctions required to write to 'other' IQ")
243 .flags(total);
244
245 wbPenalizedRate
246 .name(name() + ".WB:penalized_rate")
247 .desc ("fraction of instructions written-back that wrote to 'other' IQ")
248 .flags(total);
249
250 wbPenalizedRate = wbPenalized / writebackCount;
251
252 wbFanout
253 .name(name() + ".WB:fanout")
254 .desc("average fanout of values written-back")
255 .flags(total);
256
257 wbFanout = producerInst / consumerInst;
258
259 wbRate
260 .name(name() + ".WB:rate")
261 .desc("insts written-back per cycle")
262 .flags(total);
263 wbRate = writebackCount / cpu->numCycles;
264}
265
266template<class Impl>
267void
268DefaultIEW<Impl>::initStage()
269{
270 for (int tid=0; tid < numThreads; tid++) {
271 toRename->iewInfo[tid].usedIQ = true;
272 toRename->iewInfo[tid].freeIQEntries =
273 instQueue.numFreeEntries(tid);
274
275 toRename->iewInfo[tid].usedLSQ = true;
276 toRename->iewInfo[tid].freeLSQEntries =
277 ldstQueue.numFreeEntries(tid);
278 }
279}
280
281template<class Impl>
282void
283DefaultIEW<Impl>::setCPU(O3CPU *cpu_ptr)
284{
285 DPRINTF(IEW, "Setting CPU pointer.\n");
286 cpu = cpu_ptr;
287
288 instQueue.setCPU(cpu_ptr);
289 ldstQueue.setCPU(cpu_ptr);
290
291 cpu->activateStage(O3CPU::IEWIdx);
292}
293
294template<class Impl>
295void
296DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
297{
298 DPRINTF(IEW, "Setting time buffer pointer.\n");
299 timeBuffer = tb_ptr;
300
301 // Setup wire to read information from time buffer, from commit.
302 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
303
304 // Setup wire to write information back to previous stages.
305 toRename = timeBuffer->getWire(0);
306
307 toFetch = timeBuffer->getWire(0);
308
309 // Instruction queue also needs main time buffer.
310 instQueue.setTimeBuffer(tb_ptr);
311}
312
313template<class Impl>
314void
315DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
316{
317 DPRINTF(IEW, "Setting rename queue pointer.\n");
318 renameQueue = rq_ptr;
319
320 // Setup wire to read information from rename queue.
321 fromRename = renameQueue->getWire(-renameToIEWDelay);
322}
323
324template<class Impl>
325void
326DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
327{
328 DPRINTF(IEW, "Setting IEW queue pointer.\n");
329 iewQueue = iq_ptr;
330
331 // Setup wire to write instructions to commit.
332 toCommit = iewQueue->getWire(0);
333}
334
335template<class Impl>
336void
337DefaultIEW<Impl>::setActiveThreads(std::list<unsigned> *at_ptr)
338{
339 DPRINTF(IEW, "Setting active threads list pointer.\n");
340 activeThreads = at_ptr;
341
342 ldstQueue.setActiveThreads(at_ptr);
343 instQueue.setActiveThreads(at_ptr);
344}
345
346template<class Impl>
347void
348DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
349{
350 DPRINTF(IEW, "Setting scoreboard pointer.\n");
351 scoreboard = sb_ptr;
352}
353
354template <class Impl>
355bool
356DefaultIEW<Impl>::drain()
357{
358 // IEW is ready to drain at any time.
359 cpu->signalDrained();
360 return true;
361}
362
363template <class Impl>
364void
365DefaultIEW<Impl>::resume()
366{
367}
368
369template <class Impl>
370void
371DefaultIEW<Impl>::switchOut()
372{
373 // Clear any state.
374 switchedOut = true;
375 assert(insts[0].empty());
376 assert(skidBuffer[0].empty());
377
378 instQueue.switchOut();
379 ldstQueue.switchOut();
380 fuPool->switchOut();
381
382 for (int i = 0; i < numThreads; i++) {
383 while (!insts[i].empty())
384 insts[i].pop();
385 while (!skidBuffer[i].empty())
386 skidBuffer[i].pop();
387 }
388}
389
390template <class Impl>
391void
392DefaultIEW<Impl>::takeOverFrom()
393{
394 // Reset all state.
395 _status = Active;
396 exeStatus = Running;
397 wbStatus = Idle;
398 switchedOut = false;
399
400 instQueue.takeOverFrom();
401 ldstQueue.takeOverFrom();
402 fuPool->takeOverFrom();
403
404 initStage();
405 cpu->activityThisCycle();
406
407 for (int i=0; i < numThreads; i++) {
408 dispatchStatus[i] = Running;
409 stalls[i].commit = false;
410 fetchRedirect[i] = false;
411 }
412
413 updateLSQNextCycle = false;
414
415 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
416 issueToExecQueue.advance();
417 }
418}
419
420template<class Impl>
421void
422DefaultIEW<Impl>::squash(unsigned tid)
423{
424 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n",
425 tid);
426
427 // Tell the IQ to start squashing.
428 instQueue.squash(tid);
429
430 // Tell the LDSTQ to start squashing.
431#if ISA_HAS_DELAY_SLOT
432 ldstQueue.squash(fromCommit->commitInfo[tid].bdelayDoneSeqNum, tid);
433#else
434 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
435#endif
436 updatedQueues = true;
437
438 // Clear the skid buffer in case it has any data in it.
439 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
440 tid, fromCommit->commitInfo[tid].bdelayDoneSeqNum);
441
442 while (!skidBuffer[tid].empty()) {
443#if ISA_HAS_DELAY_SLOT
444 if (skidBuffer[tid].front()->seqNum <=
445 fromCommit->commitInfo[tid].bdelayDoneSeqNum) {
446 DPRINTF(IEW, "[tid:%i]: Cannot remove skidbuffer instructions "
447 "that occur before delay slot [sn:%i].\n",
448 fromCommit->commitInfo[tid].bdelayDoneSeqNum,
449 tid);
450 break;
451 } else {
452 DPRINTF(IEW, "[tid:%i]: Removing instruction [sn:%i] from "
453 "skidBuffer.\n", tid, skidBuffer[tid].front()->seqNum);
454 }
455#endif
456 if (skidBuffer[tid].front()->isLoad() ||
457 skidBuffer[tid].front()->isStore() ) {
458 toRename->iewInfo[tid].dispatchedToLSQ++;
459 }
460
461 toRename->iewInfo[tid].dispatched++;
462
463 skidBuffer[tid].pop();
464 }
465
466 bdelayDoneSeqNum[tid] = fromCommit->commitInfo[tid].bdelayDoneSeqNum;
467
468 emptyRenameInsts(tid);
469}
470
471template<class Impl>
472void
473DefaultIEW<Impl>::squashDueToBranch(DynInstPtr &inst, unsigned tid)
474{
475 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %#x "
476 "[sn:%i].\n", tid, inst->readPC(), inst->seqNum);
477
478 toCommit->squash[tid] = true;
479 toCommit->squashedSeqNum[tid] = inst->seqNum;
480 toCommit->mispredPC[tid] = inst->readPC();
481 toCommit->branchMispredict[tid] = true;
482
483#if ISA_HAS_DELAY_SLOT
484 bool branch_taken = inst->readNextNPC() !=
485 (inst->readNextPC() + sizeof(TheISA::MachInst));
486
487 toCommit->branchTaken[tid] = branch_taken;
488
489 toCommit->condDelaySlotBranch[tid] = inst->isCondDelaySlot();
490
491 if (inst->isCondDelaySlot() && branch_taken) {
492 toCommit->nextPC[tid] = inst->readNextPC();
493 } else {
494 toCommit->nextPC[tid] = inst->readNextNPC();
495 }
496#else
497 toCommit->branchTaken[tid] = inst->readNextPC() !=
498 (inst->readPC() + sizeof(TheISA::MachInst));
499 toCommit->nextPC[tid] = inst->readNextPC();
500#endif
501
502 toCommit->includeSquashInst[tid] = false;
503
504 wroteToTimeBuffer = true;
505}
506
507template<class Impl>
508void
509DefaultIEW<Impl>::squashDueToMemOrder(DynInstPtr &inst, unsigned tid)
510{
511 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, "
512 "PC: %#x [sn:%i].\n", tid, inst->readPC(), inst->seqNum);
513
514 toCommit->squash[tid] = true;
515 toCommit->squashedSeqNum[tid] = inst->seqNum;
516 toCommit->nextPC[tid] = inst->readNextPC();
517
518 toCommit->includeSquashInst[tid] = false;
519
520 wroteToTimeBuffer = true;
521}
522
523template<class Impl>
524void
525DefaultIEW<Impl>::squashDueToMemBlocked(DynInstPtr &inst, unsigned tid)
526{
527 DPRINTF(IEW, "[tid:%i]: Memory blocked, squashing load and younger insts, "
528 "PC: %#x [sn:%i].\n", tid, inst->readPC(), inst->seqNum);
529
530 toCommit->squash[tid] = true;
531 toCommit->squashedSeqNum[tid] = inst->seqNum;
532 toCommit->nextPC[tid] = inst->readPC();
533
534 // Must include the broadcasted SN in the squash.
535 toCommit->includeSquashInst[tid] = true;
536
537 ldstQueue.setLoadBlockedHandled(tid);
538
539 wroteToTimeBuffer = true;
540}
541
542template<class Impl>
543void
544DefaultIEW<Impl>::block(unsigned tid)
545{
546 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
547
548 if (dispatchStatus[tid] != Blocked &&
549 dispatchStatus[tid] != Unblocking) {
550 toRename->iewBlock[tid] = true;
551 wroteToTimeBuffer = true;
552 }
553
554 // Add the current inputs to the skid buffer so they can be
555 // reprocessed when this stage unblocks.
556 skidInsert(tid);
557
558 dispatchStatus[tid] = Blocked;
559}
560
561template<class Impl>
562void
563DefaultIEW<Impl>::unblock(unsigned tid)
564{
565 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
566 "buffer %u.\n",tid, tid);
567
568 // If the skid bufffer is empty, signal back to previous stages to unblock.
569 // Also switch status to running.
570 if (skidBuffer[tid].empty()) {
571 toRename->iewUnblock[tid] = true;
572 wroteToTimeBuffer = true;
573 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
574 dispatchStatus[tid] = Running;
575 }
576}
577
578template<class Impl>
579void
580DefaultIEW<Impl>::wakeDependents(DynInstPtr &inst)
581{
582 instQueue.wakeDependents(inst);
583}
584
585template<class Impl>
586void
587DefaultIEW<Impl>::rescheduleMemInst(DynInstPtr &inst)
588{
589 instQueue.rescheduleMemInst(inst);
590}
591
592template<class Impl>
593void
594DefaultIEW<Impl>::replayMemInst(DynInstPtr &inst)
595{
596 instQueue.replayMemInst(inst);
597}
598
599template<class Impl>
600void
601DefaultIEW<Impl>::instToCommit(DynInstPtr &inst)
602{
603 // First check the time slot that this instruction will write
604 // to. If there are free write ports at the time, then go ahead
605 // and write the instruction to that time. If there are not,
606 // keep looking back to see where's the first time there's a
607 // free slot.
608 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
609 ++wbNumInst;
610 if (wbNumInst == wbWidth) {
611 ++wbCycle;
612 wbNumInst = 0;
613 }
614
615 assert((wbCycle * wbWidth + wbNumInst) <= wbMax);
616 }
617
618 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
619 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
620 // Add finished instruction to queue to commit.
621 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
622 (*iewQueue)[wbCycle].size++;
623}
624
625template <class Impl>
626unsigned
627DefaultIEW<Impl>::validInstsFromRename()
628{
629 unsigned inst_count = 0;
630
631 for (int i=0; i<fromRename->size; i++) {
632 if (!fromRename->insts[i]->isSquashed())
633 inst_count++;
634 }
635
636 return inst_count;
637}
638
639template<class Impl>
640void
641DefaultIEW<Impl>::skidInsert(unsigned tid)
642{
643 DynInstPtr inst = NULL;
644
645 while (!insts[tid].empty()) {
646 inst = insts[tid].front();
647
648 insts[tid].pop();
649
650 DPRINTF(Decode,"[tid:%i]: Inserting [sn:%lli] PC:%#x into "
651 "dispatch skidBuffer %i\n",tid, inst->seqNum,
652 inst->readPC(),tid);
653
654 skidBuffer[tid].push(inst);
655 }
656
657 assert(skidBuffer[tid].size() <= skidBufferMax &&
658 "Skidbuffer Exceeded Max Size");
659}
660
661template<class Impl>
662int
663DefaultIEW<Impl>::skidCount()
664{
665 int max=0;
666
667 std::list<unsigned>::iterator threads = (*activeThreads).begin();
668
669 while (threads != (*activeThreads).end()) {
670 unsigned thread_count = skidBuffer[*threads++].size();
671 if (max < thread_count)
672 max = thread_count;
673 }
674
675 return max;
676}
677
678template<class Impl>
679bool
680DefaultIEW<Impl>::skidsEmpty()
681{
682 std::list<unsigned>::iterator threads = (*activeThreads).begin();
683
684 while (threads != (*activeThreads).end()) {
685 if (!skidBuffer[*threads++].empty())
686 return false;
687 }
688
689 return true;
690}
691
692template <class Impl>
693void
694DefaultIEW<Impl>::updateStatus()
695{
696 bool any_unblocking = false;
697
698 std::list<unsigned>::iterator threads = (*activeThreads).begin();
699
700 threads = (*activeThreads).begin();
701
702 while (threads != (*activeThreads).end()) {
703 unsigned tid = *threads++;
704
705 if (dispatchStatus[tid] == Unblocking) {
706 any_unblocking = true;
707 break;
708 }
709 }
710
711 // If there are no ready instructions waiting to be scheduled by the IQ,
712 // and there's no stores waiting to write back, and dispatch is not
713 // unblocking, then there is no internal activity for the IEW stage.
714 if (_status == Active && !instQueue.hasReadyInsts() &&
715 !ldstQueue.willWB() && !any_unblocking) {
716 DPRINTF(IEW, "IEW switching to idle\n");
717
718 deactivateStage();
719
720 _status = Inactive;
721 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
722 ldstQueue.willWB() ||
723 any_unblocking)) {
724 // Otherwise there is internal activity. Set to active.
725 DPRINTF(IEW, "IEW switching to active\n");
726
727 activateStage();
728
729 _status = Active;
730 }
731}
732
733template <class Impl>
734void
735DefaultIEW<Impl>::resetEntries()
736{
737 instQueue.resetEntries();
738 ldstQueue.resetEntries();
739}
740
741template <class Impl>
742void
743DefaultIEW<Impl>::readStallSignals(unsigned tid)
744{
745 if (fromCommit->commitBlock[tid]) {
746 stalls[tid].commit = true;
747 }
748
749 if (fromCommit->commitUnblock[tid]) {
750 assert(stalls[tid].commit);
751 stalls[tid].commit = false;
752 }
753}
754
755template <class Impl>
756bool
757DefaultIEW<Impl>::checkStall(unsigned tid)
758{
759 bool ret_val(false);
760
761 if (stalls[tid].commit) {
762 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
763 ret_val = true;
764 } else if (instQueue.isFull(tid)) {
765 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
766 ret_val = true;
767 } else if (ldstQueue.isFull(tid)) {
768 DPRINTF(IEW,"[tid:%i]: Stall: LSQ is full\n",tid);
769
770 if (ldstQueue.numLoads(tid) > 0 ) {
771
772 DPRINTF(IEW,"[tid:%i]: LSQ oldest load: [sn:%i] \n",
773 tid,ldstQueue.getLoadHeadSeqNum(tid));
774 }
775
776 if (ldstQueue.numStores(tid) > 0) {
777
778 DPRINTF(IEW,"[tid:%i]: LSQ oldest store: [sn:%i] \n",
779 tid,ldstQueue.getStoreHeadSeqNum(tid));
780 }
781
782 ret_val = true;
783 } else if (ldstQueue.isStalled(tid)) {
784 DPRINTF(IEW,"[tid:%i]: Stall: LSQ stall detected.\n",tid);
785 ret_val = true;
786 }
787
788 return ret_val;
789}
790
791template <class Impl>
792void
793DefaultIEW<Impl>::checkSignalsAndUpdate(unsigned tid)
794{
795 // Check if there's a squash signal, squash if there is
796 // Check stall signals, block if there is.
797 // If status was Blocked
798 // if so then go to unblocking
799 // If status was Squashing
800 // check if squashing is not high. Switch to running this cycle.
801
802 readStallSignals(tid);
803
804 if (fromCommit->commitInfo[tid].squash) {
805 squash(tid);
806
807 if (dispatchStatus[tid] == Blocked ||
808 dispatchStatus[tid] == Unblocking) {
809 toRename->iewUnblock[tid] = true;
810 wroteToTimeBuffer = true;
811 }
812
813 dispatchStatus[tid] = Squashing;
814
815 fetchRedirect[tid] = false;
816 return;
817 }
818
819 if (fromCommit->commitInfo[tid].robSquashing) {
820 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
821
822 dispatchStatus[tid] = Squashing;
823
824 emptyRenameInsts(tid);
825 wroteToTimeBuffer = true;
826 return;
827 }
828
829 if (checkStall(tid)) {
830 block(tid);
831 dispatchStatus[tid] = Blocked;
832 return;
833 }
834
835 if (dispatchStatus[tid] == Blocked) {
836 // Status from previous cycle was blocked, but there are no more stall
837 // conditions. Switch over to unblocking.
838 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
839 tid);
840
841 dispatchStatus[tid] = Unblocking;
842
843 unblock(tid);
844
845 return;
846 }
847
848 if (dispatchStatus[tid] == Squashing) {
849 // Switch status to running if rename isn't being told to block or
850 // squash this cycle.
851 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
852 tid);
853
854 dispatchStatus[tid] = Running;
855
856 return;
857 }
858}
859
860template <class Impl>
861void
862DefaultIEW<Impl>::sortInsts()
863{
864 int insts_from_rename = fromRename->size;
865#ifdef DEBUG
866#if !ISA_HAS_DELAY_SLOT
867 for (int i = 0; i < numThreads; i++)
868 assert(insts[i].empty());
869#endif
870#endif
871 for (int i = 0; i < insts_from_rename; ++i) {
872 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
873 }
874}
875
876template <class Impl>
877void
878DefaultIEW<Impl>::emptyRenameInsts(unsigned tid)
879{
880 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions until "
881 "[sn:%i].\n", tid, bdelayDoneSeqNum[tid]);
882
883 while (!insts[tid].empty()) {
884#if ISA_HAS_DELAY_SLOT
885 if (insts[tid].front()->seqNum <= bdelayDoneSeqNum[tid]) {
886 DPRINTF(IEW, "[tid:%i]: Done removing, cannot remove instruction"
887 " that occurs at or before delay slot [sn:%i].\n",
888 tid, bdelayDoneSeqNum[tid]);
889 break;
890 } else {
891 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instruction "
892 "[sn:%i].\n", tid, insts[tid].front()->seqNum);
893 }
894#endif
895
896 if (insts[tid].front()->isLoad() ||
897 insts[tid].front()->isStore() ) {
898 toRename->iewInfo[tid].dispatchedToLSQ++;
899 }
900
901 toRename->iewInfo[tid].dispatched++;
902
903 insts[tid].pop();
904 }
905}
906
907template <class Impl>
908void
909DefaultIEW<Impl>::emptyRenameInsts(unsigned tid)
910{
911 while (!insts[tid].empty()) {
912 if (insts[tid].front()->isLoad() ||
913 insts[tid].front()->isStore() ) {
914 toRename->iewInfo[tid].dispatchedToLSQ++;
915 }
916
917 toRename->iewInfo[tid].dispatched++;
918
919 insts[tid].pop();
920 }
921}
922
923template <class Impl>
924void
925DefaultIEW<Impl>::wakeCPU()
926{
927 cpu->wakeCPU();
928}
929
930template <class Impl>
931void
932DefaultIEW<Impl>::activityThisCycle()
933{
934 DPRINTF(Activity, "Activity this cycle.\n");
935 cpu->activityThisCycle();
936}
937
938template <class Impl>
939inline void
940DefaultIEW<Impl>::activateStage()
941{
942 DPRINTF(Activity, "Activating stage.\n");
943 cpu->activateStage(O3CPU::IEWIdx);
944}
945
946template <class Impl>
947inline void
948DefaultIEW<Impl>::deactivateStage()
949{
950 DPRINTF(Activity, "Deactivating stage.\n");
951 cpu->deactivateStage(O3CPU::IEWIdx);
952}
953
954template<class Impl>
955void
956DefaultIEW<Impl>::dispatch(unsigned tid)
957{
958 // If status is Running or idle,
959 // call dispatchInsts()
960 // If status is Unblocking,
961 // buffer any instructions coming from rename
962 // continue trying to empty skid buffer
963 // check if stall conditions have passed
964
965 if (dispatchStatus[tid] == Blocked) {
966 ++iewBlockCycles;
967
968 } else if (dispatchStatus[tid] == Squashing) {
969 ++iewSquashCycles;
970 }
971
972 // Dispatch should try to dispatch as many instructions as its bandwidth
973 // will allow, as long as it is not currently blocked.
974 if (dispatchStatus[tid] == Running ||
975 dispatchStatus[tid] == Idle) {
976 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
977 "dispatch.\n", tid);
978
979 dispatchInsts(tid);
980 } else if (dispatchStatus[tid] == Unblocking) {
981 // Make sure that the skid buffer has something in it if the
982 // status is unblocking.
983 assert(!skidsEmpty());
984
985 // If the status was unblocking, then instructions from the skid
986 // buffer were used. Remove those instructions and handle
987 // the rest of unblocking.
988 dispatchInsts(tid);
989
990 ++iewUnblockCycles;
991
992 if (validInstsFromRename() && dispatchedAllInsts) {
993 // Add the current inputs to the skid buffer so they can be
994 // reprocessed when this stage unblocks.
995 skidInsert(tid);
996 }
997
998 unblock(tid);
999 }
1000}
1001
1002template <class Impl>
1003void
1004DefaultIEW<Impl>::dispatchInsts(unsigned tid)
1005{
1006 dispatchedAllInsts = true;
1007
1008 // Obtain instructions from skid buffer if unblocking, or queue from rename
1009 // otherwise.
1010 std::queue<DynInstPtr> &insts_to_dispatch =
1011 dispatchStatus[tid] == Unblocking ?
1012 skidBuffer[tid] : insts[tid];
1013
1014 int insts_to_add = insts_to_dispatch.size();
1015
1016 DynInstPtr inst;
1017 bool add_to_iq = false;
1018 int dis_num_inst = 0;
1019
1020 // Loop through the instructions, putting them in the instruction
1021 // queue.
1022 for ( ; dis_num_inst < insts_to_add &&
1023 dis_num_inst < dispatchWidth;
1024 ++dis_num_inst)
1025 {
1026 inst = insts_to_dispatch.front();
1027
1028 if (dispatchStatus[tid] == Unblocking) {
1029 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
1030 "buffer\n", tid);
1031 }
1032
1033 // Make sure there's a valid instruction there.
1034 assert(inst);
1035
1036 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %#x [sn:%lli] [tid:%i] to "
1037 "IQ.\n",
1038 tid, inst->readPC(), inst->seqNum, inst->threadNumber);
1039
1040 // Be sure to mark these instructions as ready so that the
1041 // commit stage can go ahead and execute them, and mark
1042 // them as issued so the IQ doesn't reprocess them.
1043
1044 // Check for squashed instructions.
1045 if (inst->isSquashed()) {
1046 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
1047 "not adding to IQ.\n", tid);
1048
1049 ++iewDispSquashedInsts;
1050
1051 insts_to_dispatch.pop();
1052
1053 //Tell Rename That An Instruction has been processed
1054 if (inst->isLoad() || inst->isStore()) {
1055 toRename->iewInfo[tid].dispatchedToLSQ++;
1056 }
1057 toRename->iewInfo[tid].dispatched++;
1058
1059 continue;
1060 }
1061
1062 // Check for full conditions.
1063 if (instQueue.isFull(tid)) {
1064 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1065
1066 // Call function to start blocking.
1067 block(tid);
1068
1069 // Set unblock to false. Special case where we are using
1070 // skidbuffer (unblocking) instructions but then we still
1071 // get full in the IQ.
1072 toRename->iewUnblock[tid] = false;
1073
1074 dispatchedAllInsts = false;
1075
1076 ++iewIQFullEvents;
1077 break;
1078 } else if (ldstQueue.isFull(tid)) {
1079 DPRINTF(IEW, "[tid:%i]: Issue: LSQ has become full.\n",tid);
1080
1081 // Call function to start blocking.
1082 block(tid);
1083
1084 // Set unblock to false. Special case where we are using
1085 // skidbuffer (unblocking) instructions but then we still
1086 // get full in the IQ.
1087 toRename->iewUnblock[tid] = false;
1088
1089 dispatchedAllInsts = false;
1090
1091 ++iewLSQFullEvents;
1092 break;
1093 }
1094
1095 // Otherwise issue the instruction just fine.
1096 if (inst->isLoad()) {
1097 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1098 "encountered, adding to LSQ.\n", tid);
1099
1100 // Reserve a spot in the load store queue for this
1101 // memory access.
1102 ldstQueue.insertLoad(inst);
1103
1104 ++iewDispLoadInsts;
1105
1106 add_to_iq = true;
1107
1108 toRename->iewInfo[tid].dispatchedToLSQ++;
1109 } else if (inst->isStore()) {
1110 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1111 "encountered, adding to LSQ.\n", tid);
1112
1113 ldstQueue.insertStore(inst);
1114
1115 ++iewDispStoreInsts;
1116
1117 if (inst->isStoreConditional()) {
1118 // Store conditionals need to be set as "canCommit()"
1119 // so that commit can process them when they reach the
1120 // head of commit.
1121 // @todo: This is somewhat specific to Alpha.
1122 inst->setCanCommit();
1123 instQueue.insertNonSpec(inst);
1124 add_to_iq = false;
1125
1126 ++iewDispNonSpecInsts;
1127 } else {
1128 add_to_iq = true;
1129 }
1130
1131 toRename->iewInfo[tid].dispatchedToLSQ++;
1132#if FULL_SYSTEM
1133 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1134 // Same as non-speculative stores.
1135 inst->setCanCommit();
1136 instQueue.insertBarrier(inst);
1137 add_to_iq = false;
1138#endif
1139 } else if (inst->isNonSpeculative()) {
1140 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1141 "encountered, skipping.\n", tid);
1142
1143 // Same as non-speculative stores.
1144 inst->setCanCommit();
1145
1146 // Specifically insert it as nonspeculative.
1147 instQueue.insertNonSpec(inst);
1148
1149 ++iewDispNonSpecInsts;
1150
1151 add_to_iq = false;
1152 } else if (inst->isNop()) {
1153 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1154 "skipping.\n", tid);
1155
1156 inst->setIssued();
1157 inst->setExecuted();
1158 inst->setCanCommit();
1159
1160 instQueue.recordProducer(inst);
1161
1162 iewExecutedNop[tid]++;
1163
1164 add_to_iq = false;
1165 } else if (inst->isExecuted()) {
1166 assert(0 && "Instruction shouldn't be executed.\n");
1167 DPRINTF(IEW, "Issue: Executed branch encountered, "
1168 "skipping.\n");
1169
1170 inst->setIssued();
1171 inst->setCanCommit();
1172
1173 instQueue.recordProducer(inst);
1174
1175 add_to_iq = false;
1176 } else {
1177 add_to_iq = true;
1178 }
1179
1180 // If the instruction queue is not full, then add the
1181 // instruction.
1182 if (add_to_iq) {
1183 instQueue.insert(inst);
1184 }
1185
1186 insts_to_dispatch.pop();
1187
1188 toRename->iewInfo[tid].dispatched++;
1189
1190 ++iewDispatchedInsts;
1191 }
1192
1193 if (!insts_to_dispatch.empty()) {
1194 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1195 block(tid);
1196 toRename->iewUnblock[tid] = false;
1197 }
1198
1199 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1200 dispatchStatus[tid] = Running;
1201
1202 updatedQueues = true;
1203 }
1204
1205 dis_num_inst = 0;
1206}
1207
1208template <class Impl>
1209void
1210DefaultIEW<Impl>::printAvailableInsts()
1211{
1212 int inst = 0;
1213
1214 std::cout << "Available Instructions: ";
1215
1216 while (fromIssue->insts[inst]) {
1217
1218 if (inst%3==0) std::cout << "\n\t";
1219
1220 std::cout << "PC: " << fromIssue->insts[inst]->readPC()
1221 << " TN: " << fromIssue->insts[inst]->threadNumber
1222 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1223
1224 inst++;
1225
1226 }
1227
1228 std::cout << "\n";
1229}
1230
1231template <class Impl>
1232void
1233DefaultIEW<Impl>::executeInsts()
1234{
1235 wbNumInst = 0;
1236 wbCycle = 0;
1237
1238 std::list<unsigned>::iterator threads = (*activeThreads).begin();
1239
1240 while (threads != (*activeThreads).end()) {
1241 unsigned tid = *threads++;
1242 fetchRedirect[tid] = false;
1243 }
1244
1245 // Uncomment this if you want to see all available instructions.
1246// printAvailableInsts();
1247
1248 // Execute/writeback any instructions that are available.
1249 int insts_to_execute = fromIssue->size;
1250 int inst_num = 0;
1251 for (; inst_num < insts_to_execute;
1252 ++inst_num) {
1253
1254 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1255
1256 DynInstPtr inst = instQueue.getInstToExecute();
1257
1258 DPRINTF(IEW, "Execute: Processing PC %#x, [tid:%i] [sn:%i].\n",
1259 inst->readPC(), inst->threadNumber,inst->seqNum);
1260
1261 // Check if the instruction is squashed; if so then skip it
1262 if (inst->isSquashed()) {
1263 DPRINTF(IEW, "Execute: Instruction was squashed.\n");
1264
1265 // Consider this instruction executed so that commit can go
1266 // ahead and retire the instruction.
1267 inst->setExecuted();
1268
1269 // Not sure if I should set this here or just let commit try to
1270 // commit any squashed instructions. I like the latter a bit more.
1271 inst->setCanCommit();
1272
1273 ++iewExecSquashedInsts;
1274
1275 decrWb(inst->seqNum);
1276 continue;
1277 }
1278
1279 Fault fault = NoFault;
1280
1281 // Execute instruction.
1282 // Note that if the instruction faults, it will be handled
1283 // at the commit stage.
1284 if (inst->isMemRef() &&
1285 (!inst->isDataPrefetch() && !inst->isInstPrefetch())) {
1286 DPRINTF(IEW, "Execute: Calculating address for memory "
1287 "reference.\n");
1288
1289 // Tell the LDSTQ to execute this instruction (if it is a load).
1290 if (inst->isLoad()) {
1291 // Loads will mark themselves as executed, and their writeback
1292 // event adds the instruction to the queue to commit
1293 fault = ldstQueue.executeLoad(inst);
1294 } else if (inst->isStore()) {
1295 fault = ldstQueue.executeStore(inst);
1296
1297 // If the store had a fault then it may not have a mem req
1298 if (!inst->isStoreConditional() && fault == NoFault) {
1299 inst->setExecuted();
1300
1301 instToCommit(inst);
1302 } else if (fault != NoFault) {
1303 // If the instruction faulted, then we need to send it along to commit
1304 // without the instruction completing.
1305
1306 // Send this instruction to commit, also make sure iew stage
1307 // realizes there is activity.
1308 inst->setExecuted();
1309
1310 instToCommit(inst);
1311 activityThisCycle();
1312 }
1313
1314 // Store conditionals will mark themselves as
1315 // executed, and their writeback event will add the
1316 // instruction to the queue to commit.
1317 } else {
1318 panic("Unexpected memory type!\n");
1319 }
1320
1321 } else {
1322 inst->execute();
1323
1324 inst->setExecuted();
1325
1326 instToCommit(inst);
1327 }
1328
1329 updateExeInstStats(inst);
1330
1331 // Check if branch prediction was correct, if not then we need
1332 // to tell commit to squash in flight instructions. Only
1333 // handle this if there hasn't already been something that
1334 // redirects fetch in this group of instructions.
1335
1336 // This probably needs to prioritize the redirects if a different
1337 // scheduler is used. Currently the scheduler schedules the oldest
1338 // instruction first, so the branch resolution order will be correct.
1339 unsigned tid = inst->threadNumber;
1340
1341 if (!fetchRedirect[tid]) {
1342
1343 if (inst->mispredicted()) {
1344 fetchRedirect[tid] = true;
1345
1346 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1347#if ISA_HAS_DELAY_SLOT
1348 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x.\n",
1349 inst->nextNPC);
1350#else
1351 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x.\n",
1352 inst->nextPC);
1353#endif
1354 // If incorrect, then signal the ROB that it must be squashed.
1355 squashDueToBranch(inst, tid);
1356
1357 if (inst->predTaken()) {
1358 predictedTakenIncorrect++;
1359 } else {
1360 predictedNotTakenIncorrect++;
1361 }
1362 } else if (ldstQueue.violation(tid)) {
1363 fetchRedirect[tid] = true;
1364
1365 // If there was an ordering violation, then get the
1366 // DynInst that caused the violation. Note that this
1367 // clears the violation signal.
1368 DynInstPtr violator;
1369 violator = ldstQueue.getMemDepViolator(tid);
1370
1371 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1372 "%#x, inst PC: %#x. Addr is: %#x.\n",
1373 violator->readPC(), inst->readPC(), inst->physEffAddr);
1374
1375 // Tell the instruction queue that a violation has occured.
1376 instQueue.violation(inst, violator);
1377
1378 // Squash.
1379 squashDueToMemOrder(inst,tid);
1380
1381 ++memOrderViolationEvents;
1382 } else if (ldstQueue.loadBlocked(tid) &&
1383 !ldstQueue.isLoadBlockedHandled(tid)) {
1384 fetchRedirect[tid] = true;
1385
1386 DPRINTF(IEW, "Load operation couldn't execute because the "
1387 "memory system is blocked. PC: %#x [sn:%lli]\n",
1388 inst->readPC(), inst->seqNum);
1389
1390 squashDueToMemBlocked(inst, tid);
1391 }
1392 }
1393 }
1394
1395 // Update and record activity if we processed any instructions.
1396 if (inst_num) {
1397 if (exeStatus == Idle) {
1398 exeStatus = Running;
1399 }
1400
1401 updatedQueues = true;
1402
1403 cpu->activityThisCycle();
1404 }
1405
1406 // Need to reset this in case a writeback event needs to write into the
1407 // iew queue. That way the writeback event will write into the correct
1408 // spot in the queue.
1409 wbNumInst = 0;
1410}
1411
1412template <class Impl>
1413void
1414DefaultIEW<Impl>::writebackInsts()
1415{
1416 // Loop through the head of the time buffer and wake any
1417 // dependents. These instructions are about to write back. Also
1418 // mark scoreboard that this instruction is finally complete.
1419 // Either have IEW have direct access to scoreboard, or have this
1420 // as part of backwards communication.
1421 for (int inst_num = 0; inst_num < issueWidth &&
1422 toCommit->insts[inst_num]; inst_num++) {
1423 DynInstPtr inst = toCommit->insts[inst_num];
1424 int tid = inst->threadNumber;
1425
1426 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %#x.\n",
1427 inst->seqNum, inst->readPC());
1428
1429 iewInstsToCommit[tid]++;
1430
1431 // Some instructions will be sent to commit without having
1432 // executed because they need commit to handle them.
1433 // E.g. Uncached loads have not actually executed when they
1434 // are first sent to commit. Instead commit must tell the LSQ
1435 // when it's ready to execute the uncached load.
1436 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1437 int dependents = instQueue.wakeDependents(inst);
1438
1439 for (int i = 0; i < inst->numDestRegs(); i++) {
1440 //mark as Ready
1441 DPRINTF(IEW,"Setting Destination Register %i\n",
1442 inst->renamedDestRegIdx(i));
1443 scoreboard->setReg(inst->renamedDestRegIdx(i));
1444 }
1445
1446 if (dependents) {
1447 producerInst[tid]++;
1448 consumerInst[tid]+= dependents;
1449 }
1450 writebackCount[tid]++;
1451 }
1452
1453 decrWb(inst->seqNum);
1454 }
1455}
1456
1457template<class Impl>
1458void
1459DefaultIEW<Impl>::tick()
1460{
1461 wbNumInst = 0;
1462 wbCycle = 0;
1463
1464 wroteToTimeBuffer = false;
1465 updatedQueues = false;
1466
1467 sortInsts();
1468
1469 // Free function units marked as being freed this cycle.
1470 fuPool->processFreeUnits();
1471
1472 std::list<unsigned>::iterator threads = (*activeThreads).begin();
1473
1474 // Check stall and squash signals, dispatch any instructions.
1475 while (threads != (*activeThreads).end()) {
1476 unsigned tid = *threads++;
1477
1478 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1479
1480 checkSignalsAndUpdate(tid);
1481 dispatch(tid);
1482 }
1483
1484 if (exeStatus != Squashing) {
1485 executeInsts();
1486
1487 writebackInsts();
1488
1489 // Have the instruction queue try to schedule any ready instructions.
1490 // (In actuality, this scheduling is for instructions that will
1491 // be executed next cycle.)
1492 instQueue.scheduleReadyInsts();
1493
1494 // Also should advance its own time buffers if the stage ran.
1495 // Not the best place for it, but this works (hopefully).
1496 issueToExecQueue.advance();
1497 }
1498
1499 bool broadcast_free_entries = false;
1500
1501 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1502 exeStatus = Idle;
1503 updateLSQNextCycle = false;
1504
1505 broadcast_free_entries = true;
1506 }
1507
1508 // Writeback any stores using any leftover bandwidth.
1509 ldstQueue.writebackStores();
1510
1511 // Check the committed load/store signals to see if there's a load
1512 // or store to commit. Also check if it's being told to execute a
1513 // nonspeculative instruction.
1514 // This is pretty inefficient...
1515
1516 threads = (*activeThreads).begin();
1517 while (threads != (*activeThreads).end()) {
1518 unsigned tid = (*threads++);
1519
1520 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1521
1522 // Update structures based on instructions committed.
1523 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1524 !fromCommit->commitInfo[tid].squash &&
1525 !fromCommit->commitInfo[tid].robSquashing) {
1526
1527 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1528
1529 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1530
1531 updateLSQNextCycle = true;
1532 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1533 }
1534
1535 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1536
1537 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1538 if (fromCommit->commitInfo[tid].uncached) {
1539 instQueue.replayMemInst(fromCommit->commitInfo[tid].uncachedLoad);
1540 } else {
1541 instQueue.scheduleNonSpec(
1542 fromCommit->commitInfo[tid].nonSpecSeqNum);
1543 }
1544 }
1545
1546 if (broadcast_free_entries) {
1547 toFetch->iewInfo[tid].iqCount =
1548 instQueue.getCount(tid);
1549 toFetch->iewInfo[tid].ldstqCount =
1550 ldstQueue.getCount(tid);
1551
1552 toRename->iewInfo[tid].usedIQ = true;
1553 toRename->iewInfo[tid].freeIQEntries =
1554 instQueue.numFreeEntries();
1555 toRename->iewInfo[tid].usedLSQ = true;
1556 toRename->iewInfo[tid].freeLSQEntries =
1557 ldstQueue.numFreeEntries(tid);
1558
1559 wroteToTimeBuffer = true;
1560 }
1561
1562 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1563 tid, toRename->iewInfo[tid].dispatched);
1564 }
1565
1566 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1567 "LSQ has %i free entries.\n",
1568 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1569 ldstQueue.numFreeEntries());
1570
1571 updateStatus();
1572
1573 if (wroteToTimeBuffer) {
1574 DPRINTF(Activity, "Activity this cycle.\n");
1575 cpu->activityThisCycle();
1576 }
1577}
1578
1579template <class Impl>
1580void
1581DefaultIEW<Impl>::updateExeInstStats(DynInstPtr &inst)
1582{
1583 int thread_number = inst->threadNumber;
1584
1585 //
1586 // Pick off the software prefetches
1587 //
1588#ifdef TARGET_ALPHA
1589 if (inst->isDataPrefetch())
1590 iewExecutedSwp[thread_number]++;
1591 else
1592 iewIewExecutedcutedInsts++;
1593#else
1594 iewExecutedInsts++;
1595#endif
1596
1597 //
1598 // Control operations
1599 //
1600 if (inst->isControl())
1601 iewExecutedBranches[thread_number]++;
1602
1603 //
1604 // Memory operations
1605 //
1606 if (inst->isMemRef()) {
1607 iewExecutedRefs[thread_number]++;
1608
1609 if (inst->isLoad()) {
1610 iewExecLoadInsts[thread_number]++;
1611 }
1612 }
1613}
909DefaultIEW<Impl>::wakeCPU()
910{
911 cpu->wakeCPU();
912}
913
914template <class Impl>
915void
916DefaultIEW<Impl>::activityThisCycle()
917{
918 DPRINTF(Activity, "Activity this cycle.\n");
919 cpu->activityThisCycle();
920}
921
922template <class Impl>
923inline void
924DefaultIEW<Impl>::activateStage()
925{
926 DPRINTF(Activity, "Activating stage.\n");
927 cpu->activateStage(O3CPU::IEWIdx);
928}
929
930template <class Impl>
931inline void
932DefaultIEW<Impl>::deactivateStage()
933{
934 DPRINTF(Activity, "Deactivating stage.\n");
935 cpu->deactivateStage(O3CPU::IEWIdx);
936}
937
938template<class Impl>
939void
940DefaultIEW<Impl>::dispatch(unsigned tid)
941{
942 // If status is Running or idle,
943 // call dispatchInsts()
944 // If status is Unblocking,
945 // buffer any instructions coming from rename
946 // continue trying to empty skid buffer
947 // check if stall conditions have passed
948
949 if (dispatchStatus[tid] == Blocked) {
950 ++iewBlockCycles;
951
952 } else if (dispatchStatus[tid] == Squashing) {
953 ++iewSquashCycles;
954 }
955
956 // Dispatch should try to dispatch as many instructions as its bandwidth
957 // will allow, as long as it is not currently blocked.
958 if (dispatchStatus[tid] == Running ||
959 dispatchStatus[tid] == Idle) {
960 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
961 "dispatch.\n", tid);
962
963 dispatchInsts(tid);
964 } else if (dispatchStatus[tid] == Unblocking) {
965 // Make sure that the skid buffer has something in it if the
966 // status is unblocking.
967 assert(!skidsEmpty());
968
969 // If the status was unblocking, then instructions from the skid
970 // buffer were used. Remove those instructions and handle
971 // the rest of unblocking.
972 dispatchInsts(tid);
973
974 ++iewUnblockCycles;
975
976 if (validInstsFromRename() && dispatchedAllInsts) {
977 // Add the current inputs to the skid buffer so they can be
978 // reprocessed when this stage unblocks.
979 skidInsert(tid);
980 }
981
982 unblock(tid);
983 }
984}
985
986template <class Impl>
987void
988DefaultIEW<Impl>::dispatchInsts(unsigned tid)
989{
990 dispatchedAllInsts = true;
991
992 // Obtain instructions from skid buffer if unblocking, or queue from rename
993 // otherwise.
994 std::queue<DynInstPtr> &insts_to_dispatch =
995 dispatchStatus[tid] == Unblocking ?
996 skidBuffer[tid] : insts[tid];
997
998 int insts_to_add = insts_to_dispatch.size();
999
1000 DynInstPtr inst;
1001 bool add_to_iq = false;
1002 int dis_num_inst = 0;
1003
1004 // Loop through the instructions, putting them in the instruction
1005 // queue.
1006 for ( ; dis_num_inst < insts_to_add &&
1007 dis_num_inst < dispatchWidth;
1008 ++dis_num_inst)
1009 {
1010 inst = insts_to_dispatch.front();
1011
1012 if (dispatchStatus[tid] == Unblocking) {
1013 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
1014 "buffer\n", tid);
1015 }
1016
1017 // Make sure there's a valid instruction there.
1018 assert(inst);
1019
1020 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %#x [sn:%lli] [tid:%i] to "
1021 "IQ.\n",
1022 tid, inst->readPC(), inst->seqNum, inst->threadNumber);
1023
1024 // Be sure to mark these instructions as ready so that the
1025 // commit stage can go ahead and execute them, and mark
1026 // them as issued so the IQ doesn't reprocess them.
1027
1028 // Check for squashed instructions.
1029 if (inst->isSquashed()) {
1030 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
1031 "not adding to IQ.\n", tid);
1032
1033 ++iewDispSquashedInsts;
1034
1035 insts_to_dispatch.pop();
1036
1037 //Tell Rename That An Instruction has been processed
1038 if (inst->isLoad() || inst->isStore()) {
1039 toRename->iewInfo[tid].dispatchedToLSQ++;
1040 }
1041 toRename->iewInfo[tid].dispatched++;
1042
1043 continue;
1044 }
1045
1046 // Check for full conditions.
1047 if (instQueue.isFull(tid)) {
1048 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1049
1050 // Call function to start blocking.
1051 block(tid);
1052
1053 // Set unblock to false. Special case where we are using
1054 // skidbuffer (unblocking) instructions but then we still
1055 // get full in the IQ.
1056 toRename->iewUnblock[tid] = false;
1057
1058 dispatchedAllInsts = false;
1059
1060 ++iewIQFullEvents;
1061 break;
1062 } else if (ldstQueue.isFull(tid)) {
1063 DPRINTF(IEW, "[tid:%i]: Issue: LSQ has become full.\n",tid);
1064
1065 // Call function to start blocking.
1066 block(tid);
1067
1068 // Set unblock to false. Special case where we are using
1069 // skidbuffer (unblocking) instructions but then we still
1070 // get full in the IQ.
1071 toRename->iewUnblock[tid] = false;
1072
1073 dispatchedAllInsts = false;
1074
1075 ++iewLSQFullEvents;
1076 break;
1077 }
1078
1079 // Otherwise issue the instruction just fine.
1080 if (inst->isLoad()) {
1081 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1082 "encountered, adding to LSQ.\n", tid);
1083
1084 // Reserve a spot in the load store queue for this
1085 // memory access.
1086 ldstQueue.insertLoad(inst);
1087
1088 ++iewDispLoadInsts;
1089
1090 add_to_iq = true;
1091
1092 toRename->iewInfo[tid].dispatchedToLSQ++;
1093 } else if (inst->isStore()) {
1094 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1095 "encountered, adding to LSQ.\n", tid);
1096
1097 ldstQueue.insertStore(inst);
1098
1099 ++iewDispStoreInsts;
1100
1101 if (inst->isStoreConditional()) {
1102 // Store conditionals need to be set as "canCommit()"
1103 // so that commit can process them when they reach the
1104 // head of commit.
1105 // @todo: This is somewhat specific to Alpha.
1106 inst->setCanCommit();
1107 instQueue.insertNonSpec(inst);
1108 add_to_iq = false;
1109
1110 ++iewDispNonSpecInsts;
1111 } else {
1112 add_to_iq = true;
1113 }
1114
1115 toRename->iewInfo[tid].dispatchedToLSQ++;
1116#if FULL_SYSTEM
1117 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1118 // Same as non-speculative stores.
1119 inst->setCanCommit();
1120 instQueue.insertBarrier(inst);
1121 add_to_iq = false;
1122#endif
1123 } else if (inst->isNonSpeculative()) {
1124 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1125 "encountered, skipping.\n", tid);
1126
1127 // Same as non-speculative stores.
1128 inst->setCanCommit();
1129
1130 // Specifically insert it as nonspeculative.
1131 instQueue.insertNonSpec(inst);
1132
1133 ++iewDispNonSpecInsts;
1134
1135 add_to_iq = false;
1136 } else if (inst->isNop()) {
1137 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1138 "skipping.\n", tid);
1139
1140 inst->setIssued();
1141 inst->setExecuted();
1142 inst->setCanCommit();
1143
1144 instQueue.recordProducer(inst);
1145
1146 iewExecutedNop[tid]++;
1147
1148 add_to_iq = false;
1149 } else if (inst->isExecuted()) {
1150 assert(0 && "Instruction shouldn't be executed.\n");
1151 DPRINTF(IEW, "Issue: Executed branch encountered, "
1152 "skipping.\n");
1153
1154 inst->setIssued();
1155 inst->setCanCommit();
1156
1157 instQueue.recordProducer(inst);
1158
1159 add_to_iq = false;
1160 } else {
1161 add_to_iq = true;
1162 }
1163
1164 // If the instruction queue is not full, then add the
1165 // instruction.
1166 if (add_to_iq) {
1167 instQueue.insert(inst);
1168 }
1169
1170 insts_to_dispatch.pop();
1171
1172 toRename->iewInfo[tid].dispatched++;
1173
1174 ++iewDispatchedInsts;
1175 }
1176
1177 if (!insts_to_dispatch.empty()) {
1178 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1179 block(tid);
1180 toRename->iewUnblock[tid] = false;
1181 }
1182
1183 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1184 dispatchStatus[tid] = Running;
1185
1186 updatedQueues = true;
1187 }
1188
1189 dis_num_inst = 0;
1190}
1191
1192template <class Impl>
1193void
1194DefaultIEW<Impl>::printAvailableInsts()
1195{
1196 int inst = 0;
1197
1198 std::cout << "Available Instructions: ";
1199
1200 while (fromIssue->insts[inst]) {
1201
1202 if (inst%3==0) std::cout << "\n\t";
1203
1204 std::cout << "PC: " << fromIssue->insts[inst]->readPC()
1205 << " TN: " << fromIssue->insts[inst]->threadNumber
1206 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1207
1208 inst++;
1209
1210 }
1211
1212 std::cout << "\n";
1213}
1214
1215template <class Impl>
1216void
1217DefaultIEW<Impl>::executeInsts()
1218{
1219 wbNumInst = 0;
1220 wbCycle = 0;
1221
1222 std::list<unsigned>::iterator threads = (*activeThreads).begin();
1223
1224 while (threads != (*activeThreads).end()) {
1225 unsigned tid = *threads++;
1226 fetchRedirect[tid] = false;
1227 }
1228
1229 // Uncomment this if you want to see all available instructions.
1230// printAvailableInsts();
1231
1232 // Execute/writeback any instructions that are available.
1233 int insts_to_execute = fromIssue->size;
1234 int inst_num = 0;
1235 for (; inst_num < insts_to_execute;
1236 ++inst_num) {
1237
1238 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1239
1240 DynInstPtr inst = instQueue.getInstToExecute();
1241
1242 DPRINTF(IEW, "Execute: Processing PC %#x, [tid:%i] [sn:%i].\n",
1243 inst->readPC(), inst->threadNumber,inst->seqNum);
1244
1245 // Check if the instruction is squashed; if so then skip it
1246 if (inst->isSquashed()) {
1247 DPRINTF(IEW, "Execute: Instruction was squashed.\n");
1248
1249 // Consider this instruction executed so that commit can go
1250 // ahead and retire the instruction.
1251 inst->setExecuted();
1252
1253 // Not sure if I should set this here or just let commit try to
1254 // commit any squashed instructions. I like the latter a bit more.
1255 inst->setCanCommit();
1256
1257 ++iewExecSquashedInsts;
1258
1259 decrWb(inst->seqNum);
1260 continue;
1261 }
1262
1263 Fault fault = NoFault;
1264
1265 // Execute instruction.
1266 // Note that if the instruction faults, it will be handled
1267 // at the commit stage.
1268 if (inst->isMemRef() &&
1269 (!inst->isDataPrefetch() && !inst->isInstPrefetch())) {
1270 DPRINTF(IEW, "Execute: Calculating address for memory "
1271 "reference.\n");
1272
1273 // Tell the LDSTQ to execute this instruction (if it is a load).
1274 if (inst->isLoad()) {
1275 // Loads will mark themselves as executed, and their writeback
1276 // event adds the instruction to the queue to commit
1277 fault = ldstQueue.executeLoad(inst);
1278 } else if (inst->isStore()) {
1279 fault = ldstQueue.executeStore(inst);
1280
1281 // If the store had a fault then it may not have a mem req
1282 if (!inst->isStoreConditional() && fault == NoFault) {
1283 inst->setExecuted();
1284
1285 instToCommit(inst);
1286 } else if (fault != NoFault) {
1287 // If the instruction faulted, then we need to send it along to commit
1288 // without the instruction completing.
1289
1290 // Send this instruction to commit, also make sure iew stage
1291 // realizes there is activity.
1292 inst->setExecuted();
1293
1294 instToCommit(inst);
1295 activityThisCycle();
1296 }
1297
1298 // Store conditionals will mark themselves as
1299 // executed, and their writeback event will add the
1300 // instruction to the queue to commit.
1301 } else {
1302 panic("Unexpected memory type!\n");
1303 }
1304
1305 } else {
1306 inst->execute();
1307
1308 inst->setExecuted();
1309
1310 instToCommit(inst);
1311 }
1312
1313 updateExeInstStats(inst);
1314
1315 // Check if branch prediction was correct, if not then we need
1316 // to tell commit to squash in flight instructions. Only
1317 // handle this if there hasn't already been something that
1318 // redirects fetch in this group of instructions.
1319
1320 // This probably needs to prioritize the redirects if a different
1321 // scheduler is used. Currently the scheduler schedules the oldest
1322 // instruction first, so the branch resolution order will be correct.
1323 unsigned tid = inst->threadNumber;
1324
1325 if (!fetchRedirect[tid]) {
1326
1327 if (inst->mispredicted()) {
1328 fetchRedirect[tid] = true;
1329
1330 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1331#if ISA_HAS_DELAY_SLOT
1332 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x.\n",
1333 inst->nextNPC);
1334#else
1335 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x.\n",
1336 inst->nextPC);
1337#endif
1338 // If incorrect, then signal the ROB that it must be squashed.
1339 squashDueToBranch(inst, tid);
1340
1341 if (inst->predTaken()) {
1342 predictedTakenIncorrect++;
1343 } else {
1344 predictedNotTakenIncorrect++;
1345 }
1346 } else if (ldstQueue.violation(tid)) {
1347 fetchRedirect[tid] = true;
1348
1349 // If there was an ordering violation, then get the
1350 // DynInst that caused the violation. Note that this
1351 // clears the violation signal.
1352 DynInstPtr violator;
1353 violator = ldstQueue.getMemDepViolator(tid);
1354
1355 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1356 "%#x, inst PC: %#x. Addr is: %#x.\n",
1357 violator->readPC(), inst->readPC(), inst->physEffAddr);
1358
1359 // Tell the instruction queue that a violation has occured.
1360 instQueue.violation(inst, violator);
1361
1362 // Squash.
1363 squashDueToMemOrder(inst,tid);
1364
1365 ++memOrderViolationEvents;
1366 } else if (ldstQueue.loadBlocked(tid) &&
1367 !ldstQueue.isLoadBlockedHandled(tid)) {
1368 fetchRedirect[tid] = true;
1369
1370 DPRINTF(IEW, "Load operation couldn't execute because the "
1371 "memory system is blocked. PC: %#x [sn:%lli]\n",
1372 inst->readPC(), inst->seqNum);
1373
1374 squashDueToMemBlocked(inst, tid);
1375 }
1376 }
1377 }
1378
1379 // Update and record activity if we processed any instructions.
1380 if (inst_num) {
1381 if (exeStatus == Idle) {
1382 exeStatus = Running;
1383 }
1384
1385 updatedQueues = true;
1386
1387 cpu->activityThisCycle();
1388 }
1389
1390 // Need to reset this in case a writeback event needs to write into the
1391 // iew queue. That way the writeback event will write into the correct
1392 // spot in the queue.
1393 wbNumInst = 0;
1394}
1395
1396template <class Impl>
1397void
1398DefaultIEW<Impl>::writebackInsts()
1399{
1400 // Loop through the head of the time buffer and wake any
1401 // dependents. These instructions are about to write back. Also
1402 // mark scoreboard that this instruction is finally complete.
1403 // Either have IEW have direct access to scoreboard, or have this
1404 // as part of backwards communication.
1405 for (int inst_num = 0; inst_num < issueWidth &&
1406 toCommit->insts[inst_num]; inst_num++) {
1407 DynInstPtr inst = toCommit->insts[inst_num];
1408 int tid = inst->threadNumber;
1409
1410 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %#x.\n",
1411 inst->seqNum, inst->readPC());
1412
1413 iewInstsToCommit[tid]++;
1414
1415 // Some instructions will be sent to commit without having
1416 // executed because they need commit to handle them.
1417 // E.g. Uncached loads have not actually executed when they
1418 // are first sent to commit. Instead commit must tell the LSQ
1419 // when it's ready to execute the uncached load.
1420 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1421 int dependents = instQueue.wakeDependents(inst);
1422
1423 for (int i = 0; i < inst->numDestRegs(); i++) {
1424 //mark as Ready
1425 DPRINTF(IEW,"Setting Destination Register %i\n",
1426 inst->renamedDestRegIdx(i));
1427 scoreboard->setReg(inst->renamedDestRegIdx(i));
1428 }
1429
1430 if (dependents) {
1431 producerInst[tid]++;
1432 consumerInst[tid]+= dependents;
1433 }
1434 writebackCount[tid]++;
1435 }
1436
1437 decrWb(inst->seqNum);
1438 }
1439}
1440
1441template<class Impl>
1442void
1443DefaultIEW<Impl>::tick()
1444{
1445 wbNumInst = 0;
1446 wbCycle = 0;
1447
1448 wroteToTimeBuffer = false;
1449 updatedQueues = false;
1450
1451 sortInsts();
1452
1453 // Free function units marked as being freed this cycle.
1454 fuPool->processFreeUnits();
1455
1456 std::list<unsigned>::iterator threads = (*activeThreads).begin();
1457
1458 // Check stall and squash signals, dispatch any instructions.
1459 while (threads != (*activeThreads).end()) {
1460 unsigned tid = *threads++;
1461
1462 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1463
1464 checkSignalsAndUpdate(tid);
1465 dispatch(tid);
1466 }
1467
1468 if (exeStatus != Squashing) {
1469 executeInsts();
1470
1471 writebackInsts();
1472
1473 // Have the instruction queue try to schedule any ready instructions.
1474 // (In actuality, this scheduling is for instructions that will
1475 // be executed next cycle.)
1476 instQueue.scheduleReadyInsts();
1477
1478 // Also should advance its own time buffers if the stage ran.
1479 // Not the best place for it, but this works (hopefully).
1480 issueToExecQueue.advance();
1481 }
1482
1483 bool broadcast_free_entries = false;
1484
1485 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1486 exeStatus = Idle;
1487 updateLSQNextCycle = false;
1488
1489 broadcast_free_entries = true;
1490 }
1491
1492 // Writeback any stores using any leftover bandwidth.
1493 ldstQueue.writebackStores();
1494
1495 // Check the committed load/store signals to see if there's a load
1496 // or store to commit. Also check if it's being told to execute a
1497 // nonspeculative instruction.
1498 // This is pretty inefficient...
1499
1500 threads = (*activeThreads).begin();
1501 while (threads != (*activeThreads).end()) {
1502 unsigned tid = (*threads++);
1503
1504 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1505
1506 // Update structures based on instructions committed.
1507 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1508 !fromCommit->commitInfo[tid].squash &&
1509 !fromCommit->commitInfo[tid].robSquashing) {
1510
1511 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1512
1513 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1514
1515 updateLSQNextCycle = true;
1516 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1517 }
1518
1519 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1520
1521 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1522 if (fromCommit->commitInfo[tid].uncached) {
1523 instQueue.replayMemInst(fromCommit->commitInfo[tid].uncachedLoad);
1524 } else {
1525 instQueue.scheduleNonSpec(
1526 fromCommit->commitInfo[tid].nonSpecSeqNum);
1527 }
1528 }
1529
1530 if (broadcast_free_entries) {
1531 toFetch->iewInfo[tid].iqCount =
1532 instQueue.getCount(tid);
1533 toFetch->iewInfo[tid].ldstqCount =
1534 ldstQueue.getCount(tid);
1535
1536 toRename->iewInfo[tid].usedIQ = true;
1537 toRename->iewInfo[tid].freeIQEntries =
1538 instQueue.numFreeEntries();
1539 toRename->iewInfo[tid].usedLSQ = true;
1540 toRename->iewInfo[tid].freeLSQEntries =
1541 ldstQueue.numFreeEntries(tid);
1542
1543 wroteToTimeBuffer = true;
1544 }
1545
1546 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1547 tid, toRename->iewInfo[tid].dispatched);
1548 }
1549
1550 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1551 "LSQ has %i free entries.\n",
1552 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1553 ldstQueue.numFreeEntries());
1554
1555 updateStatus();
1556
1557 if (wroteToTimeBuffer) {
1558 DPRINTF(Activity, "Activity this cycle.\n");
1559 cpu->activityThisCycle();
1560 }
1561}
1562
1563template <class Impl>
1564void
1565DefaultIEW<Impl>::updateExeInstStats(DynInstPtr &inst)
1566{
1567 int thread_number = inst->threadNumber;
1568
1569 //
1570 // Pick off the software prefetches
1571 //
1572#ifdef TARGET_ALPHA
1573 if (inst->isDataPrefetch())
1574 iewExecutedSwp[thread_number]++;
1575 else
1576 iewIewExecutedcutedInsts++;
1577#else
1578 iewExecutedInsts++;
1579#endif
1580
1581 //
1582 // Control operations
1583 //
1584 if (inst->isControl())
1585 iewExecutedBranches[thread_number]++;
1586
1587 //
1588 // Memory operations
1589 //
1590 if (inst->isMemRef()) {
1591 iewExecutedRefs[thread_number]++;
1592
1593 if (inst->isLoad()) {
1594 iewExecLoadInsts[thread_number]++;
1595 }
1596 }
1597}