1/*
2 * Copyright (c) 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#include "arch/types.hh"
44#include "base/trace.hh"
45#include "config/the_isa.hh"
46#include "cpu/o3/decode.hh"
47#include "cpu/inst_seq.hh"
48#include "debug/Activity.hh"
49#include "debug/Decode.hh"
| 1/*
2 * Copyright (c) 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#include "arch/types.hh"
44#include "base/trace.hh"
45#include "config/the_isa.hh"
46#include "cpu/o3/decode.hh"
47#include "cpu/inst_seq.hh"
48#include "debug/Activity.hh"
49#include "debug/Decode.hh"
|
50#include "params/DerivO3CPU.hh"
51#include "sim/full_system.hh"
52
53// clang complains about std::set being overloaded with Packet::set if
54// we open up the entire namespace std
55using std::list;
56
57template<class Impl>
58DefaultDecode<Impl>::DefaultDecode(O3CPU *_cpu, DerivO3CPUParams *params)
59 : cpu(_cpu),
60 renameToDecodeDelay(params->renameToDecodeDelay),
61 iewToDecodeDelay(params->iewToDecodeDelay),
62 commitToDecodeDelay(params->commitToDecodeDelay),
63 fetchToDecodeDelay(params->fetchToDecodeDelay),
64 decodeWidth(params->decodeWidth),
65 numThreads(params->numThreads)
66{
67 // @todo: Make into a parameter
68 skidBufferMax = (fetchToDecodeDelay + 1) * params->fetchWidth;
69}
70
71template<class Impl>
72void
73DefaultDecode<Impl>::startupStage()
74{
75 resetStage();
76}
77
78template<class Impl>
79void
80DefaultDecode<Impl>::resetStage()
81{
82 _status = Inactive;
83
84 // Setup status, make sure stall signals are clear.
85 for (ThreadID tid = 0; tid < numThreads; ++tid) {
86 decodeStatus[tid] = Idle;
87
88 stalls[tid].rename = false;
89 stalls[tid].iew = false;
90 stalls[tid].commit = false;
91 }
92}
93
94template <class Impl>
95std::string
96DefaultDecode<Impl>::name() const
97{
98 return cpu->name() + ".decode";
99}
100
101template <class Impl>
102void
103DefaultDecode<Impl>::regStats()
104{
105 decodeIdleCycles
106 .name(name() + ".IdleCycles")
107 .desc("Number of cycles decode is idle")
108 .prereq(decodeIdleCycles);
109 decodeBlockedCycles
110 .name(name() + ".BlockedCycles")
111 .desc("Number of cycles decode is blocked")
112 .prereq(decodeBlockedCycles);
113 decodeRunCycles
114 .name(name() + ".RunCycles")
115 .desc("Number of cycles decode is running")
116 .prereq(decodeRunCycles);
117 decodeUnblockCycles
118 .name(name() + ".UnblockCycles")
119 .desc("Number of cycles decode is unblocking")
120 .prereq(decodeUnblockCycles);
121 decodeSquashCycles
122 .name(name() + ".SquashCycles")
123 .desc("Number of cycles decode is squashing")
124 .prereq(decodeSquashCycles);
125 decodeBranchResolved
126 .name(name() + ".BranchResolved")
127 .desc("Number of times decode resolved a branch")
128 .prereq(decodeBranchResolved);
129 decodeBranchMispred
130 .name(name() + ".BranchMispred")
131 .desc("Number of times decode detected a branch misprediction")
132 .prereq(decodeBranchMispred);
133 decodeControlMispred
134 .name(name() + ".ControlMispred")
135 .desc("Number of times decode detected an instruction incorrectly"
136 " predicted as a control")
137 .prereq(decodeControlMispred);
138 decodeDecodedInsts
139 .name(name() + ".DecodedInsts")
140 .desc("Number of instructions handled by decode")
141 .prereq(decodeDecodedInsts);
142 decodeSquashedInsts
143 .name(name() + ".SquashedInsts")
144 .desc("Number of squashed instructions handled by decode")
145 .prereq(decodeSquashedInsts);
146}
147
148template<class Impl>
149void
150DefaultDecode<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
151{
152 timeBuffer = tb_ptr;
153
154 // Setup wire to write information back to fetch.
155 toFetch = timeBuffer->getWire(0);
156
157 // Create wires to get information from proper places in time buffer.
158 fromRename = timeBuffer->getWire(-renameToDecodeDelay);
159 fromIEW = timeBuffer->getWire(-iewToDecodeDelay);
160 fromCommit = timeBuffer->getWire(-commitToDecodeDelay);
161}
162
163template<class Impl>
164void
165DefaultDecode<Impl>::setDecodeQueue(TimeBuffer<DecodeStruct> *dq_ptr)
166{
167 decodeQueue = dq_ptr;
168
169 // Setup wire to write information to proper place in decode queue.
170 toRename = decodeQueue->getWire(0);
171}
172
173template<class Impl>
174void
175DefaultDecode<Impl>::setFetchQueue(TimeBuffer<FetchStruct> *fq_ptr)
176{
177 fetchQueue = fq_ptr;
178
179 // Setup wire to read information from fetch queue.
180 fromFetch = fetchQueue->getWire(-fetchToDecodeDelay);
181}
182
183template<class Impl>
184void
185DefaultDecode<Impl>::setActiveThreads(std::list<ThreadID> *at_ptr)
186{
187 activeThreads = at_ptr;
188}
189
190template <class Impl>
191void
192DefaultDecode<Impl>::drainSanityCheck() const
193{
194 for (ThreadID tid = 0; tid < numThreads; ++tid) {
195 assert(insts[tid].empty());
196 assert(skidBuffer[tid].empty());
197 }
198}
199
200template<class Impl>
201bool
202DefaultDecode<Impl>::checkStall(ThreadID tid) const
203{
204 bool ret_val = false;
205
206 if (stalls[tid].rename) {
207 DPRINTF(Decode,"[tid:%i]: Stall fom Rename stage detected.\n", tid);
208 ret_val = true;
209 } else if (stalls[tid].iew) {
210 DPRINTF(Decode,"[tid:%i]: Stall fom IEW stage detected.\n", tid);
211 ret_val = true;
212 } else if (stalls[tid].commit) {
213 DPRINTF(Decode,"[tid:%i]: Stall fom Commit stage detected.\n", tid);
214 ret_val = true;
215 }
216
217 return ret_val;
218}
219
220template<class Impl>
221inline bool
222DefaultDecode<Impl>::fetchInstsValid()
223{
224 return fromFetch->size > 0;
225}
226
227template<class Impl>
228bool
229DefaultDecode<Impl>::block(ThreadID tid)
230{
231 DPRINTF(Decode, "[tid:%u]: Blocking.\n", tid);
232
233 // Add the current inputs to the skid buffer so they can be
234 // reprocessed when this stage unblocks.
235 skidInsert(tid);
236
237 // If the decode status is blocked or unblocking then decode has not yet
238 // signalled fetch to unblock. In that case, there is no need to tell
239 // fetch to block.
240 if (decodeStatus[tid] != Blocked) {
241 // Set the status to Blocked.
242 decodeStatus[tid] = Blocked;
243
244 if (toFetch->decodeUnblock[tid]) {
245 toFetch->decodeUnblock[tid] = false;
246 } else {
247 toFetch->decodeBlock[tid] = true;
248 wroteToTimeBuffer = true;
249 }
250
251 return true;
252 }
253
254 return false;
255}
256
257template<class Impl>
258bool
259DefaultDecode<Impl>::unblock(ThreadID tid)
260{
261 // Decode is done unblocking only if the skid buffer is empty.
262 if (skidBuffer[tid].empty()) {
263 DPRINTF(Decode, "[tid:%u]: Done unblocking.\n", tid);
264 toFetch->decodeUnblock[tid] = true;
265 wroteToTimeBuffer = true;
266
267 decodeStatus[tid] = Running;
268 return true;
269 }
270
271 DPRINTF(Decode, "[tid:%u]: Currently unblocking.\n", tid);
272
273 return false;
274}
275
276template<class Impl>
277void
278DefaultDecode<Impl>::squash(DynInstPtr &inst, ThreadID tid)
279{
280 DPRINTF(Decode, "[tid:%i]: [sn:%i] Squashing due to incorrect branch "
281 "prediction detected at decode.\n", tid, inst->seqNum);
282
283 // Send back mispredict information.
284 toFetch->decodeInfo[tid].branchMispredict = true;
285 toFetch->decodeInfo[tid].predIncorrect = true;
286 toFetch->decodeInfo[tid].mispredictInst = inst;
287 toFetch->decodeInfo[tid].squash = true;
288 toFetch->decodeInfo[tid].doneSeqNum = inst->seqNum;
289 toFetch->decodeInfo[tid].nextPC = inst->branchTarget();
290 toFetch->decodeInfo[tid].branchTaken = inst->pcState().branching();
291 toFetch->decodeInfo[tid].squashInst = inst;
292 if (toFetch->decodeInfo[tid].mispredictInst->isUncondCtrl()) {
293 toFetch->decodeInfo[tid].branchTaken = true;
294 }
295
296 InstSeqNum squash_seq_num = inst->seqNum;
297
298 // Might have to tell fetch to unblock.
299 if (decodeStatus[tid] == Blocked ||
300 decodeStatus[tid] == Unblocking) {
301 toFetch->decodeUnblock[tid] = 1;
302 }
303
304 // Set status to squashing.
305 decodeStatus[tid] = Squashing;
306
307 for (int i=0; i<fromFetch->size; i++) {
308 if (fromFetch->insts[i]->threadNumber == tid &&
309 fromFetch->insts[i]->seqNum > squash_seq_num) {
310 fromFetch->insts[i]->setSquashed();
311 }
312 }
313
314 // Clear the instruction list and skid buffer in case they have any
315 // insts in them.
316 while (!insts[tid].empty()) {
317 insts[tid].pop();
318 }
319
320 while (!skidBuffer[tid].empty()) {
321 skidBuffer[tid].pop();
322 }
323
324 // Squash instructions up until this one
325 cpu->removeInstsUntil(squash_seq_num, tid);
326}
327
328template<class Impl>
329unsigned
330DefaultDecode<Impl>::squash(ThreadID tid)
331{
332 DPRINTF(Decode, "[tid:%i]: Squashing.\n",tid);
333
334 if (decodeStatus[tid] == Blocked ||
335 decodeStatus[tid] == Unblocking) {
336 if (FullSystem) {
337 toFetch->decodeUnblock[tid] = 1;
338 } else {
339 // In syscall emulation, we can have both a block and a squash due
340 // to a syscall in the same cycle. This would cause both signals
341 // to be high. This shouldn't happen in full system.
342 // @todo: Determine if this still happens.
343 if (toFetch->decodeBlock[tid])
344 toFetch->decodeBlock[tid] = 0;
345 else
346 toFetch->decodeUnblock[tid] = 1;
347 }
348 }
349
350 // Set status to squashing.
351 decodeStatus[tid] = Squashing;
352
353 // Go through incoming instructions from fetch and squash them.
354 unsigned squash_count = 0;
355
356 for (int i=0; i<fromFetch->size; i++) {
357 if (fromFetch->insts[i]->threadNumber == tid) {
358 fromFetch->insts[i]->setSquashed();
359 squash_count++;
360 }
361 }
362
363 // Clear the instruction list and skid buffer in case they have any
364 // insts in them.
365 while (!insts[tid].empty()) {
366 insts[tid].pop();
367 }
368
369 while (!skidBuffer[tid].empty()) {
370 skidBuffer[tid].pop();
371 }
372
373 return squash_count;
374}
375
376template<class Impl>
377void
378DefaultDecode<Impl>::skidInsert(ThreadID tid)
379{
380 DynInstPtr inst = NULL;
381
382 while (!insts[tid].empty()) {
383 inst = insts[tid].front();
384
385 insts[tid].pop();
386
387 assert(tid == inst->threadNumber);
388
389 DPRINTF(Decode,"Inserting [sn:%lli] PC: %s into decode skidBuffer %i\n",
390 inst->seqNum, inst->pcState(), inst->threadNumber);
391
392 skidBuffer[tid].push(inst);
393 }
394
395 // @todo: Eventually need to enforce this by not letting a thread
396 // fetch past its skidbuffer
397 assert(skidBuffer[tid].size() <= skidBufferMax);
398}
399
400template<class Impl>
401bool
402DefaultDecode<Impl>::skidsEmpty()
403{
404 list<ThreadID>::iterator threads = activeThreads->begin();
405 list<ThreadID>::iterator end = activeThreads->end();
406
407 while (threads != end) {
408 ThreadID tid = *threads++;
409 if (!skidBuffer[tid].empty())
410 return false;
411 }
412
413 return true;
414}
415
416template<class Impl>
417void
418DefaultDecode<Impl>::updateStatus()
419{
420 bool any_unblocking = false;
421
422 list<ThreadID>::iterator threads = activeThreads->begin();
423 list<ThreadID>::iterator end = activeThreads->end();
424
425 while (threads != end) {
426 ThreadID tid = *threads++;
427
428 if (decodeStatus[tid] == Unblocking) {
429 any_unblocking = true;
430 break;
431 }
432 }
433
434 // Decode will have activity if it's unblocking.
435 if (any_unblocking) {
436 if (_status == Inactive) {
437 _status = Active;
438
439 DPRINTF(Activity, "Activating stage.\n");
440
441 cpu->activateStage(O3CPU::DecodeIdx);
442 }
443 } else {
444 // If it's not unblocking, then decode will not have any internal
445 // activity. Switch it to inactive.
446 if (_status == Active) {
447 _status = Inactive;
448 DPRINTF(Activity, "Deactivating stage.\n");
449
450 cpu->deactivateStage(O3CPU::DecodeIdx);
451 }
452 }
453}
454
455template <class Impl>
456void
457DefaultDecode<Impl>::sortInsts()
458{
459 int insts_from_fetch = fromFetch->size;
460 for (int i = 0; i < insts_from_fetch; ++i) {
461 insts[fromFetch->insts[i]->threadNumber].push(fromFetch->insts[i]);
462 }
463}
464
465template<class Impl>
466void
467DefaultDecode<Impl>::readStallSignals(ThreadID tid)
468{
469 if (fromRename->renameBlock[tid]) {
470 stalls[tid].rename = true;
471 }
472
473 if (fromRename->renameUnblock[tid]) {
474 assert(stalls[tid].rename);
475 stalls[tid].rename = false;
476 }
477
478 if (fromIEW->iewBlock[tid]) {
479 stalls[tid].iew = true;
480 }
481
482 if (fromIEW->iewUnblock[tid]) {
483 assert(stalls[tid].iew);
484 stalls[tid].iew = false;
485 }
486
487 if (fromCommit->commitBlock[tid]) {
488 stalls[tid].commit = true;
489 }
490
491 if (fromCommit->commitUnblock[tid]) {
492 assert(stalls[tid].commit);
493 stalls[tid].commit = false;
494 }
495}
496
497template <class Impl>
498bool
499DefaultDecode<Impl>::checkSignalsAndUpdate(ThreadID tid)
500{
501 // Check if there's a squash signal, squash if there is.
502 // Check stall signals, block if necessary.
503 // If status was blocked
504 // Check if stall conditions have passed
505 // if so then go to unblocking
506 // If status was Squashing
507 // check if squashing is not high. Switch to running this cycle.
508
509 // Update the per thread stall statuses.
510 readStallSignals(tid);
511
512 // Check squash signals from commit.
513 if (fromCommit->commitInfo[tid].squash) {
514
515 DPRINTF(Decode, "[tid:%u]: Squashing instructions due to squash "
516 "from commit.\n", tid);
517
518 squash(tid);
519
520 return true;
521 }
522
523 // Check ROB squash signals from commit.
524 if (fromCommit->commitInfo[tid].robSquashing) {
525 DPRINTF(Decode, "[tid:%u]: ROB is still squashing.\n", tid);
526
527 // Continue to squash.
528 decodeStatus[tid] = Squashing;
529
530 return true;
531 }
532
533 if (checkStall(tid)) {
534 return block(tid);
535 }
536
537 if (decodeStatus[tid] == Blocked) {
538 DPRINTF(Decode, "[tid:%u]: Done blocking, switching to unblocking.\n",
539 tid);
540
541 decodeStatus[tid] = Unblocking;
542
543 unblock(tid);
544
545 return true;
546 }
547
548 if (decodeStatus[tid] == Squashing) {
549 // Switch status to running if decode isn't being told to block or
550 // squash this cycle.
551 DPRINTF(Decode, "[tid:%u]: Done squashing, switching to running.\n",
552 tid);
553
554 decodeStatus[tid] = Running;
555
556 return false;
557 }
558
559 // If we've reached this point, we have not gotten any signals that
560 // cause decode to change its status. Decode remains the same as before.
561 return false;
562}
563
564template<class Impl>
565void
566DefaultDecode<Impl>::tick()
567{
568 wroteToTimeBuffer = false;
569
570 bool status_change = false;
571
572 toRenameIndex = 0;
573
574 list<ThreadID>::iterator threads = activeThreads->begin();
575 list<ThreadID>::iterator end = activeThreads->end();
576
577 sortInsts();
578
579 //Check stall and squash signals.
580 while (threads != end) {
581 ThreadID tid = *threads++;
582
583 DPRINTF(Decode,"Processing [tid:%i]\n",tid);
584 status_change = checkSignalsAndUpdate(tid) || status_change;
585
586 decode(status_change, tid);
587 }
588
589 if (status_change) {
590 updateStatus();
591 }
592
593 if (wroteToTimeBuffer) {
594 DPRINTF(Activity, "Activity this cycle.\n");
595
596 cpu->activityThisCycle();
597 }
598}
599
600template<class Impl>
601void
602DefaultDecode<Impl>::decode(bool &status_change, ThreadID tid)
603{
604 // If status is Running or idle,
605 // call decodeInsts()
606 // If status is Unblocking,
607 // buffer any instructions coming from fetch
608 // continue trying to empty skid buffer
609 // check if stall conditions have passed
610
611 if (decodeStatus[tid] == Blocked) {
612 ++decodeBlockedCycles;
613 } else if (decodeStatus[tid] == Squashing) {
614 ++decodeSquashCycles;
615 }
616
617 // Decode should try to decode as many instructions as its bandwidth
618 // will allow, as long as it is not currently blocked.
619 if (decodeStatus[tid] == Running ||
620 decodeStatus[tid] == Idle) {
621 DPRINTF(Decode, "[tid:%u]: Not blocked, so attempting to run "
622 "stage.\n",tid);
623
624 decodeInsts(tid);
625 } else if (decodeStatus[tid] == Unblocking) {
626 // Make sure that the skid buffer has something in it if the
627 // status is unblocking.
628 assert(!skidsEmpty());
629
630 // If the status was unblocking, then instructions from the skid
631 // buffer were used. Remove those instructions and handle
632 // the rest of unblocking.
633 decodeInsts(tid);
634
635 if (fetchInstsValid()) {
636 // Add the current inputs to the skid buffer so they can be
637 // reprocessed when this stage unblocks.
638 skidInsert(tid);
639 }
640
641 status_change = unblock(tid) || status_change;
642 }
643}
644
645template <class Impl>
646void
647DefaultDecode<Impl>::decodeInsts(ThreadID tid)
648{
649 // Instructions can come either from the skid buffer or the list of
650 // instructions coming from fetch, depending on decode's status.
651 int insts_available = decodeStatus[tid] == Unblocking ?
652 skidBuffer[tid].size() : insts[tid].size();
653
654 if (insts_available == 0) {
655 DPRINTF(Decode, "[tid:%u] Nothing to do, breaking out"
656 " early.\n",tid);
657 // Should I change the status to idle?
658 ++decodeIdleCycles;
659 return;
660 } else if (decodeStatus[tid] == Unblocking) {
661 DPRINTF(Decode, "[tid:%u] Unblocking, removing insts from skid "
662 "buffer.\n",tid);
663 ++decodeUnblockCycles;
664 } else if (decodeStatus[tid] == Running) {
665 ++decodeRunCycles;
666 }
667
668 DynInstPtr inst;
669
670 std::queue<DynInstPtr>
671 &insts_to_decode = decodeStatus[tid] == Unblocking ?
672 skidBuffer[tid] : insts[tid];
673
674 DPRINTF(Decode, "[tid:%u]: Sending instruction to rename.\n",tid);
675
676 while (insts_available > 0 && toRenameIndex < decodeWidth) {
677 assert(!insts_to_decode.empty());
678
679 inst = insts_to_decode.front();
680
681 insts_to_decode.pop();
682
683 DPRINTF(Decode, "[tid:%u]: Processing instruction [sn:%lli] with "
684 "PC %s\n", tid, inst->seqNum, inst->pcState());
685
686 if (inst->isSquashed()) {
687 DPRINTF(Decode, "[tid:%u]: Instruction %i with PC %s is "
688 "squashed, skipping.\n",
689 tid, inst->seqNum, inst->pcState());
690
691 ++decodeSquashedInsts;
692
693 --insts_available;
694
695 continue;
696 }
697
698 // Also check if instructions have no source registers. Mark
699 // them as ready to issue at any time. Not sure if this check
700 // should exist here or at a later stage; however it doesn't matter
701 // too much for function correctness.
702 if (inst->numSrcRegs() == 0) {
703 inst->setCanIssue();
704 }
705
706 // This current instruction is valid, so add it into the decode
707 // queue. The next instruction may not be valid, so check to
708 // see if branches were predicted correctly.
709 toRename->insts[toRenameIndex] = inst;
710
711 ++(toRename->size);
712 ++toRenameIndex;
713 ++decodeDecodedInsts;
714 --insts_available;
715
716#if TRACING_ON
| 51#include "params/DerivO3CPU.hh"
52#include "sim/full_system.hh"
53
54// clang complains about std::set being overloaded with Packet::set if
55// we open up the entire namespace std
56using std::list;
57
58template<class Impl>
59DefaultDecode<Impl>::DefaultDecode(O3CPU *_cpu, DerivO3CPUParams *params)
60 : cpu(_cpu),
61 renameToDecodeDelay(params->renameToDecodeDelay),
62 iewToDecodeDelay(params->iewToDecodeDelay),
63 commitToDecodeDelay(params->commitToDecodeDelay),
64 fetchToDecodeDelay(params->fetchToDecodeDelay),
65 decodeWidth(params->decodeWidth),
66 numThreads(params->numThreads)
67{
68 // @todo: Make into a parameter
69 skidBufferMax = (fetchToDecodeDelay + 1) * params->fetchWidth;
70}
71
72template<class Impl>
73void
74DefaultDecode<Impl>::startupStage()
75{
76 resetStage();
77}
78
79template<class Impl>
80void
81DefaultDecode<Impl>::resetStage()
82{
83 _status = Inactive;
84
85 // Setup status, make sure stall signals are clear.
86 for (ThreadID tid = 0; tid < numThreads; ++tid) {
87 decodeStatus[tid] = Idle;
88
89 stalls[tid].rename = false;
90 stalls[tid].iew = false;
91 stalls[tid].commit = false;
92 }
93}
94
95template <class Impl>
96std::string
97DefaultDecode<Impl>::name() const
98{
99 return cpu->name() + ".decode";
100}
101
102template <class Impl>
103void
104DefaultDecode<Impl>::regStats()
105{
106 decodeIdleCycles
107 .name(name() + ".IdleCycles")
108 .desc("Number of cycles decode is idle")
109 .prereq(decodeIdleCycles);
110 decodeBlockedCycles
111 .name(name() + ".BlockedCycles")
112 .desc("Number of cycles decode is blocked")
113 .prereq(decodeBlockedCycles);
114 decodeRunCycles
115 .name(name() + ".RunCycles")
116 .desc("Number of cycles decode is running")
117 .prereq(decodeRunCycles);
118 decodeUnblockCycles
119 .name(name() + ".UnblockCycles")
120 .desc("Number of cycles decode is unblocking")
121 .prereq(decodeUnblockCycles);
122 decodeSquashCycles
123 .name(name() + ".SquashCycles")
124 .desc("Number of cycles decode is squashing")
125 .prereq(decodeSquashCycles);
126 decodeBranchResolved
127 .name(name() + ".BranchResolved")
128 .desc("Number of times decode resolved a branch")
129 .prereq(decodeBranchResolved);
130 decodeBranchMispred
131 .name(name() + ".BranchMispred")
132 .desc("Number of times decode detected a branch misprediction")
133 .prereq(decodeBranchMispred);
134 decodeControlMispred
135 .name(name() + ".ControlMispred")
136 .desc("Number of times decode detected an instruction incorrectly"
137 " predicted as a control")
138 .prereq(decodeControlMispred);
139 decodeDecodedInsts
140 .name(name() + ".DecodedInsts")
141 .desc("Number of instructions handled by decode")
142 .prereq(decodeDecodedInsts);
143 decodeSquashedInsts
144 .name(name() + ".SquashedInsts")
145 .desc("Number of squashed instructions handled by decode")
146 .prereq(decodeSquashedInsts);
147}
148
149template<class Impl>
150void
151DefaultDecode<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
152{
153 timeBuffer = tb_ptr;
154
155 // Setup wire to write information back to fetch.
156 toFetch = timeBuffer->getWire(0);
157
158 // Create wires to get information from proper places in time buffer.
159 fromRename = timeBuffer->getWire(-renameToDecodeDelay);
160 fromIEW = timeBuffer->getWire(-iewToDecodeDelay);
161 fromCommit = timeBuffer->getWire(-commitToDecodeDelay);
162}
163
164template<class Impl>
165void
166DefaultDecode<Impl>::setDecodeQueue(TimeBuffer<DecodeStruct> *dq_ptr)
167{
168 decodeQueue = dq_ptr;
169
170 // Setup wire to write information to proper place in decode queue.
171 toRename = decodeQueue->getWire(0);
172}
173
174template<class Impl>
175void
176DefaultDecode<Impl>::setFetchQueue(TimeBuffer<FetchStruct> *fq_ptr)
177{
178 fetchQueue = fq_ptr;
179
180 // Setup wire to read information from fetch queue.
181 fromFetch = fetchQueue->getWire(-fetchToDecodeDelay);
182}
183
184template<class Impl>
185void
186DefaultDecode<Impl>::setActiveThreads(std::list<ThreadID> *at_ptr)
187{
188 activeThreads = at_ptr;
189}
190
191template <class Impl>
192void
193DefaultDecode<Impl>::drainSanityCheck() const
194{
195 for (ThreadID tid = 0; tid < numThreads; ++tid) {
196 assert(insts[tid].empty());
197 assert(skidBuffer[tid].empty());
198 }
199}
200
201template<class Impl>
202bool
203DefaultDecode<Impl>::checkStall(ThreadID tid) const
204{
205 bool ret_val = false;
206
207 if (stalls[tid].rename) {
208 DPRINTF(Decode,"[tid:%i]: Stall fom Rename stage detected.\n", tid);
209 ret_val = true;
210 } else if (stalls[tid].iew) {
211 DPRINTF(Decode,"[tid:%i]: Stall fom IEW stage detected.\n", tid);
212 ret_val = true;
213 } else if (stalls[tid].commit) {
214 DPRINTF(Decode,"[tid:%i]: Stall fom Commit stage detected.\n", tid);
215 ret_val = true;
216 }
217
218 return ret_val;
219}
220
221template<class Impl>
222inline bool
223DefaultDecode<Impl>::fetchInstsValid()
224{
225 return fromFetch->size > 0;
226}
227
228template<class Impl>
229bool
230DefaultDecode<Impl>::block(ThreadID tid)
231{
232 DPRINTF(Decode, "[tid:%u]: Blocking.\n", tid);
233
234 // Add the current inputs to the skid buffer so they can be
235 // reprocessed when this stage unblocks.
236 skidInsert(tid);
237
238 // If the decode status is blocked or unblocking then decode has not yet
239 // signalled fetch to unblock. In that case, there is no need to tell
240 // fetch to block.
241 if (decodeStatus[tid] != Blocked) {
242 // Set the status to Blocked.
243 decodeStatus[tid] = Blocked;
244
245 if (toFetch->decodeUnblock[tid]) {
246 toFetch->decodeUnblock[tid] = false;
247 } else {
248 toFetch->decodeBlock[tid] = true;
249 wroteToTimeBuffer = true;
250 }
251
252 return true;
253 }
254
255 return false;
256}
257
258template<class Impl>
259bool
260DefaultDecode<Impl>::unblock(ThreadID tid)
261{
262 // Decode is done unblocking only if the skid buffer is empty.
263 if (skidBuffer[tid].empty()) {
264 DPRINTF(Decode, "[tid:%u]: Done unblocking.\n", tid);
265 toFetch->decodeUnblock[tid] = true;
266 wroteToTimeBuffer = true;
267
268 decodeStatus[tid] = Running;
269 return true;
270 }
271
272 DPRINTF(Decode, "[tid:%u]: Currently unblocking.\n", tid);
273
274 return false;
275}
276
277template<class Impl>
278void
279DefaultDecode<Impl>::squash(DynInstPtr &inst, ThreadID tid)
280{
281 DPRINTF(Decode, "[tid:%i]: [sn:%i] Squashing due to incorrect branch "
282 "prediction detected at decode.\n", tid, inst->seqNum);
283
284 // Send back mispredict information.
285 toFetch->decodeInfo[tid].branchMispredict = true;
286 toFetch->decodeInfo[tid].predIncorrect = true;
287 toFetch->decodeInfo[tid].mispredictInst = inst;
288 toFetch->decodeInfo[tid].squash = true;
289 toFetch->decodeInfo[tid].doneSeqNum = inst->seqNum;
290 toFetch->decodeInfo[tid].nextPC = inst->branchTarget();
291 toFetch->decodeInfo[tid].branchTaken = inst->pcState().branching();
292 toFetch->decodeInfo[tid].squashInst = inst;
293 if (toFetch->decodeInfo[tid].mispredictInst->isUncondCtrl()) {
294 toFetch->decodeInfo[tid].branchTaken = true;
295 }
296
297 InstSeqNum squash_seq_num = inst->seqNum;
298
299 // Might have to tell fetch to unblock.
300 if (decodeStatus[tid] == Blocked ||
301 decodeStatus[tid] == Unblocking) {
302 toFetch->decodeUnblock[tid] = 1;
303 }
304
305 // Set status to squashing.
306 decodeStatus[tid] = Squashing;
307
308 for (int i=0; i<fromFetch->size; i++) {
309 if (fromFetch->insts[i]->threadNumber == tid &&
310 fromFetch->insts[i]->seqNum > squash_seq_num) {
311 fromFetch->insts[i]->setSquashed();
312 }
313 }
314
315 // Clear the instruction list and skid buffer in case they have any
316 // insts in them.
317 while (!insts[tid].empty()) {
318 insts[tid].pop();
319 }
320
321 while (!skidBuffer[tid].empty()) {
322 skidBuffer[tid].pop();
323 }
324
325 // Squash instructions up until this one
326 cpu->removeInstsUntil(squash_seq_num, tid);
327}
328
329template<class Impl>
330unsigned
331DefaultDecode<Impl>::squash(ThreadID tid)
332{
333 DPRINTF(Decode, "[tid:%i]: Squashing.\n",tid);
334
335 if (decodeStatus[tid] == Blocked ||
336 decodeStatus[tid] == Unblocking) {
337 if (FullSystem) {
338 toFetch->decodeUnblock[tid] = 1;
339 } else {
340 // In syscall emulation, we can have both a block and a squash due
341 // to a syscall in the same cycle. This would cause both signals
342 // to be high. This shouldn't happen in full system.
343 // @todo: Determine if this still happens.
344 if (toFetch->decodeBlock[tid])
345 toFetch->decodeBlock[tid] = 0;
346 else
347 toFetch->decodeUnblock[tid] = 1;
348 }
349 }
350
351 // Set status to squashing.
352 decodeStatus[tid] = Squashing;
353
354 // Go through incoming instructions from fetch and squash them.
355 unsigned squash_count = 0;
356
357 for (int i=0; i<fromFetch->size; i++) {
358 if (fromFetch->insts[i]->threadNumber == tid) {
359 fromFetch->insts[i]->setSquashed();
360 squash_count++;
361 }
362 }
363
364 // Clear the instruction list and skid buffer in case they have any
365 // insts in them.
366 while (!insts[tid].empty()) {
367 insts[tid].pop();
368 }
369
370 while (!skidBuffer[tid].empty()) {
371 skidBuffer[tid].pop();
372 }
373
374 return squash_count;
375}
376
377template<class Impl>
378void
379DefaultDecode<Impl>::skidInsert(ThreadID tid)
380{
381 DynInstPtr inst = NULL;
382
383 while (!insts[tid].empty()) {
384 inst = insts[tid].front();
385
386 insts[tid].pop();
387
388 assert(tid == inst->threadNumber);
389
390 DPRINTF(Decode,"Inserting [sn:%lli] PC: %s into decode skidBuffer %i\n",
391 inst->seqNum, inst->pcState(), inst->threadNumber);
392
393 skidBuffer[tid].push(inst);
394 }
395
396 // @todo: Eventually need to enforce this by not letting a thread
397 // fetch past its skidbuffer
398 assert(skidBuffer[tid].size() <= skidBufferMax);
399}
400
401template<class Impl>
402bool
403DefaultDecode<Impl>::skidsEmpty()
404{
405 list<ThreadID>::iterator threads = activeThreads->begin();
406 list<ThreadID>::iterator end = activeThreads->end();
407
408 while (threads != end) {
409 ThreadID tid = *threads++;
410 if (!skidBuffer[tid].empty())
411 return false;
412 }
413
414 return true;
415}
416
417template<class Impl>
418void
419DefaultDecode<Impl>::updateStatus()
420{
421 bool any_unblocking = false;
422
423 list<ThreadID>::iterator threads = activeThreads->begin();
424 list<ThreadID>::iterator end = activeThreads->end();
425
426 while (threads != end) {
427 ThreadID tid = *threads++;
428
429 if (decodeStatus[tid] == Unblocking) {
430 any_unblocking = true;
431 break;
432 }
433 }
434
435 // Decode will have activity if it's unblocking.
436 if (any_unblocking) {
437 if (_status == Inactive) {
438 _status = Active;
439
440 DPRINTF(Activity, "Activating stage.\n");
441
442 cpu->activateStage(O3CPU::DecodeIdx);
443 }
444 } else {
445 // If it's not unblocking, then decode will not have any internal
446 // activity. Switch it to inactive.
447 if (_status == Active) {
448 _status = Inactive;
449 DPRINTF(Activity, "Deactivating stage.\n");
450
451 cpu->deactivateStage(O3CPU::DecodeIdx);
452 }
453 }
454}
455
456template <class Impl>
457void
458DefaultDecode<Impl>::sortInsts()
459{
460 int insts_from_fetch = fromFetch->size;
461 for (int i = 0; i < insts_from_fetch; ++i) {
462 insts[fromFetch->insts[i]->threadNumber].push(fromFetch->insts[i]);
463 }
464}
465
466template<class Impl>
467void
468DefaultDecode<Impl>::readStallSignals(ThreadID tid)
469{
470 if (fromRename->renameBlock[tid]) {
471 stalls[tid].rename = true;
472 }
473
474 if (fromRename->renameUnblock[tid]) {
475 assert(stalls[tid].rename);
476 stalls[tid].rename = false;
477 }
478
479 if (fromIEW->iewBlock[tid]) {
480 stalls[tid].iew = true;
481 }
482
483 if (fromIEW->iewUnblock[tid]) {
484 assert(stalls[tid].iew);
485 stalls[tid].iew = false;
486 }
487
488 if (fromCommit->commitBlock[tid]) {
489 stalls[tid].commit = true;
490 }
491
492 if (fromCommit->commitUnblock[tid]) {
493 assert(stalls[tid].commit);
494 stalls[tid].commit = false;
495 }
496}
497
498template <class Impl>
499bool
500DefaultDecode<Impl>::checkSignalsAndUpdate(ThreadID tid)
501{
502 // Check if there's a squash signal, squash if there is.
503 // Check stall signals, block if necessary.
504 // If status was blocked
505 // Check if stall conditions have passed
506 // if so then go to unblocking
507 // If status was Squashing
508 // check if squashing is not high. Switch to running this cycle.
509
510 // Update the per thread stall statuses.
511 readStallSignals(tid);
512
513 // Check squash signals from commit.
514 if (fromCommit->commitInfo[tid].squash) {
515
516 DPRINTF(Decode, "[tid:%u]: Squashing instructions due to squash "
517 "from commit.\n", tid);
518
519 squash(tid);
520
521 return true;
522 }
523
524 // Check ROB squash signals from commit.
525 if (fromCommit->commitInfo[tid].robSquashing) {
526 DPRINTF(Decode, "[tid:%u]: ROB is still squashing.\n", tid);
527
528 // Continue to squash.
529 decodeStatus[tid] = Squashing;
530
531 return true;
532 }
533
534 if (checkStall(tid)) {
535 return block(tid);
536 }
537
538 if (decodeStatus[tid] == Blocked) {
539 DPRINTF(Decode, "[tid:%u]: Done blocking, switching to unblocking.\n",
540 tid);
541
542 decodeStatus[tid] = Unblocking;
543
544 unblock(tid);
545
546 return true;
547 }
548
549 if (decodeStatus[tid] == Squashing) {
550 // Switch status to running if decode isn't being told to block or
551 // squash this cycle.
552 DPRINTF(Decode, "[tid:%u]: Done squashing, switching to running.\n",
553 tid);
554
555 decodeStatus[tid] = Running;
556
557 return false;
558 }
559
560 // If we've reached this point, we have not gotten any signals that
561 // cause decode to change its status. Decode remains the same as before.
562 return false;
563}
564
565template<class Impl>
566void
567DefaultDecode<Impl>::tick()
568{
569 wroteToTimeBuffer = false;
570
571 bool status_change = false;
572
573 toRenameIndex = 0;
574
575 list<ThreadID>::iterator threads = activeThreads->begin();
576 list<ThreadID>::iterator end = activeThreads->end();
577
578 sortInsts();
579
580 //Check stall and squash signals.
581 while (threads != end) {
582 ThreadID tid = *threads++;
583
584 DPRINTF(Decode,"Processing [tid:%i]\n",tid);
585 status_change = checkSignalsAndUpdate(tid) || status_change;
586
587 decode(status_change, tid);
588 }
589
590 if (status_change) {
591 updateStatus();
592 }
593
594 if (wroteToTimeBuffer) {
595 DPRINTF(Activity, "Activity this cycle.\n");
596
597 cpu->activityThisCycle();
598 }
599}
600
601template<class Impl>
602void
603DefaultDecode<Impl>::decode(bool &status_change, ThreadID tid)
604{
605 // If status is Running or idle,
606 // call decodeInsts()
607 // If status is Unblocking,
608 // buffer any instructions coming from fetch
609 // continue trying to empty skid buffer
610 // check if stall conditions have passed
611
612 if (decodeStatus[tid] == Blocked) {
613 ++decodeBlockedCycles;
614 } else if (decodeStatus[tid] == Squashing) {
615 ++decodeSquashCycles;
616 }
617
618 // Decode should try to decode as many instructions as its bandwidth
619 // will allow, as long as it is not currently blocked.
620 if (decodeStatus[tid] == Running ||
621 decodeStatus[tid] == Idle) {
622 DPRINTF(Decode, "[tid:%u]: Not blocked, so attempting to run "
623 "stage.\n",tid);
624
625 decodeInsts(tid);
626 } else if (decodeStatus[tid] == Unblocking) {
627 // Make sure that the skid buffer has something in it if the
628 // status is unblocking.
629 assert(!skidsEmpty());
630
631 // If the status was unblocking, then instructions from the skid
632 // buffer were used. Remove those instructions and handle
633 // the rest of unblocking.
634 decodeInsts(tid);
635
636 if (fetchInstsValid()) {
637 // Add the current inputs to the skid buffer so they can be
638 // reprocessed when this stage unblocks.
639 skidInsert(tid);
640 }
641
642 status_change = unblock(tid) || status_change;
643 }
644}
645
646template <class Impl>
647void
648DefaultDecode<Impl>::decodeInsts(ThreadID tid)
649{
650 // Instructions can come either from the skid buffer or the list of
651 // instructions coming from fetch, depending on decode's status.
652 int insts_available = decodeStatus[tid] == Unblocking ?
653 skidBuffer[tid].size() : insts[tid].size();
654
655 if (insts_available == 0) {
656 DPRINTF(Decode, "[tid:%u] Nothing to do, breaking out"
657 " early.\n",tid);
658 // Should I change the status to idle?
659 ++decodeIdleCycles;
660 return;
661 } else if (decodeStatus[tid] == Unblocking) {
662 DPRINTF(Decode, "[tid:%u] Unblocking, removing insts from skid "
663 "buffer.\n",tid);
664 ++decodeUnblockCycles;
665 } else if (decodeStatus[tid] == Running) {
666 ++decodeRunCycles;
667 }
668
669 DynInstPtr inst;
670
671 std::queue<DynInstPtr>
672 &insts_to_decode = decodeStatus[tid] == Unblocking ?
673 skidBuffer[tid] : insts[tid];
674
675 DPRINTF(Decode, "[tid:%u]: Sending instruction to rename.\n",tid);
676
677 while (insts_available > 0 && toRenameIndex < decodeWidth) {
678 assert(!insts_to_decode.empty());
679
680 inst = insts_to_decode.front();
681
682 insts_to_decode.pop();
683
684 DPRINTF(Decode, "[tid:%u]: Processing instruction [sn:%lli] with "
685 "PC %s\n", tid, inst->seqNum, inst->pcState());
686
687 if (inst->isSquashed()) {
688 DPRINTF(Decode, "[tid:%u]: Instruction %i with PC %s is "
689 "squashed, skipping.\n",
690 tid, inst->seqNum, inst->pcState());
691
692 ++decodeSquashedInsts;
693
694 --insts_available;
695
696 continue;
697 }
698
699 // Also check if instructions have no source registers. Mark
700 // them as ready to issue at any time. Not sure if this check
701 // should exist here or at a later stage; however it doesn't matter
702 // too much for function correctness.
703 if (inst->numSrcRegs() == 0) {
704 inst->setCanIssue();
705 }
706
707 // This current instruction is valid, so add it into the decode
708 // queue. The next instruction may not be valid, so check to
709 // see if branches were predicted correctly.
710 toRename->insts[toRenameIndex] = inst;
711
712 ++(toRename->size);
713 ++toRenameIndex;
714 ++decodeDecodedInsts;
715 --insts_available;
716
717#if TRACING_ON
|