eventq: convert all usage of events to use the new API.
[gem5.git] / src / cpu / o3 / inst_queue_impl.hh
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 * Korey Sewell
30 */
31
32 #include <limits>
33 #include <vector>
34
35 #include "cpu/o3/fu_pool.hh"
36 #include "cpu/o3/inst_queue.hh"
37 #include "enums/OpClass.hh"
38 #include "sim/core.hh"
39
40 #include "params/DerivO3CPU.hh"
41
42 template <class Impl>
43 InstructionQueue<Impl>::FUCompletion::FUCompletion(DynInstPtr &_inst,
44 int fu_idx, InstructionQueue<Impl> *iq_ptr)
45 : Event(Stat_Event_Pri), inst(_inst), fuIdx(fu_idx), iqPtr(iq_ptr),
46 freeFU(false)
47 {
48 this->setFlags(Event::AutoDelete);
49 }
50
51 template <class Impl>
52 void
53 InstructionQueue<Impl>::FUCompletion::process()
54 {
55 iqPtr->processFUCompletion(inst, freeFU ? fuIdx : -1);
56 inst = NULL;
57 }
58
59
60 template <class Impl>
61 const char *
62 InstructionQueue<Impl>::FUCompletion::description() const
63 {
64 return "Functional unit completion";
65 }
66
67 template <class Impl>
68 InstructionQueue<Impl>::InstructionQueue(O3CPU *cpu_ptr, IEW *iew_ptr,
69 DerivO3CPUParams *params)
70 : cpu(cpu_ptr),
71 iewStage(iew_ptr),
72 fuPool(params->fuPool),
73 numEntries(params->numIQEntries),
74 totalWidth(params->issueWidth),
75 numPhysIntRegs(params->numPhysIntRegs),
76 numPhysFloatRegs(params->numPhysFloatRegs),
77 commitToIEWDelay(params->commitToIEWDelay)
78 {
79 assert(fuPool);
80
81 switchedOut = false;
82
83 numThreads = params->numThreads;
84
85 // Set the number of physical registers as the number of int + float
86 numPhysRegs = numPhysIntRegs + numPhysFloatRegs;
87
88 //Create an entry for each physical register within the
89 //dependency graph.
90 dependGraph.resize(numPhysRegs);
91
92 // Resize the register scoreboard.
93 regScoreboard.resize(numPhysRegs);
94
95 //Initialize Mem Dependence Units
96 for (int i = 0; i < numThreads; i++) {
97 memDepUnit[i].init(params,i);
98 memDepUnit[i].setIQ(this);
99 }
100
101 resetState();
102
103 std::string policy = params->smtIQPolicy;
104
105 //Convert string to lowercase
106 std::transform(policy.begin(), policy.end(), policy.begin(),
107 (int(*)(int)) tolower);
108
109 //Figure out resource sharing policy
110 if (policy == "dynamic") {
111 iqPolicy = Dynamic;
112
113 //Set Max Entries to Total ROB Capacity
114 for (int i = 0; i < numThreads; i++) {
115 maxEntries[i] = numEntries;
116 }
117
118 } else if (policy == "partitioned") {
119 iqPolicy = Partitioned;
120
121 //@todo:make work if part_amt doesnt divide evenly.
122 int part_amt = numEntries / numThreads;
123
124 //Divide ROB up evenly
125 for (int i = 0; i < numThreads; i++) {
126 maxEntries[i] = part_amt;
127 }
128
129 DPRINTF(IQ, "IQ sharing policy set to Partitioned:"
130 "%i entries per thread.\n",part_amt);
131 } else if (policy == "threshold") {
132 iqPolicy = Threshold;
133
134 double threshold = (double)params->smtIQThreshold / 100;
135
136 int thresholdIQ = (int)((double)threshold * numEntries);
137
138 //Divide up by threshold amount
139 for (int i = 0; i < numThreads; i++) {
140 maxEntries[i] = thresholdIQ;
141 }
142
143 DPRINTF(IQ, "IQ sharing policy set to Threshold:"
144 "%i entries per thread.\n",thresholdIQ);
145 } else {
146 assert(0 && "Invalid IQ Sharing Policy.Options Are:{Dynamic,"
147 "Partitioned, Threshold}");
148 }
149 }
150
151 template <class Impl>
152 InstructionQueue<Impl>::~InstructionQueue()
153 {
154 dependGraph.reset();
155 #ifdef DEBUG
156 cprintf("Nodes traversed: %i, removed: %i\n",
157 dependGraph.nodesTraversed, dependGraph.nodesRemoved);
158 #endif
159 }
160
161 template <class Impl>
162 std::string
163 InstructionQueue<Impl>::name() const
164 {
165 return cpu->name() + ".iq";
166 }
167
168 template <class Impl>
169 void
170 InstructionQueue<Impl>::regStats()
171 {
172 using namespace Stats;
173 iqInstsAdded
174 .name(name() + ".iqInstsAdded")
175 .desc("Number of instructions added to the IQ (excludes non-spec)")
176 .prereq(iqInstsAdded);
177
178 iqNonSpecInstsAdded
179 .name(name() + ".iqNonSpecInstsAdded")
180 .desc("Number of non-speculative instructions added to the IQ")
181 .prereq(iqNonSpecInstsAdded);
182
183 iqInstsIssued
184 .name(name() + ".iqInstsIssued")
185 .desc("Number of instructions issued")
186 .prereq(iqInstsIssued);
187
188 iqIntInstsIssued
189 .name(name() + ".iqIntInstsIssued")
190 .desc("Number of integer instructions issued")
191 .prereq(iqIntInstsIssued);
192
193 iqFloatInstsIssued
194 .name(name() + ".iqFloatInstsIssued")
195 .desc("Number of float instructions issued")
196 .prereq(iqFloatInstsIssued);
197
198 iqBranchInstsIssued
199 .name(name() + ".iqBranchInstsIssued")
200 .desc("Number of branch instructions issued")
201 .prereq(iqBranchInstsIssued);
202
203 iqMemInstsIssued
204 .name(name() + ".iqMemInstsIssued")
205 .desc("Number of memory instructions issued")
206 .prereq(iqMemInstsIssued);
207
208 iqMiscInstsIssued
209 .name(name() + ".iqMiscInstsIssued")
210 .desc("Number of miscellaneous instructions issued")
211 .prereq(iqMiscInstsIssued);
212
213 iqSquashedInstsIssued
214 .name(name() + ".iqSquashedInstsIssued")
215 .desc("Number of squashed instructions issued")
216 .prereq(iqSquashedInstsIssued);
217
218 iqSquashedInstsExamined
219 .name(name() + ".iqSquashedInstsExamined")
220 .desc("Number of squashed instructions iterated over during squash;"
221 " mainly for profiling")
222 .prereq(iqSquashedInstsExamined);
223
224 iqSquashedOperandsExamined
225 .name(name() + ".iqSquashedOperandsExamined")
226 .desc("Number of squashed operands that are examined and possibly "
227 "removed from graph")
228 .prereq(iqSquashedOperandsExamined);
229
230 iqSquashedNonSpecRemoved
231 .name(name() + ".iqSquashedNonSpecRemoved")
232 .desc("Number of squashed non-spec instructions that were removed")
233 .prereq(iqSquashedNonSpecRemoved);
234 /*
235 queueResDist
236 .init(Num_OpClasses, 0, 99, 2)
237 .name(name() + ".IQ:residence:")
238 .desc("cycles from dispatch to issue")
239 .flags(total | pdf | cdf )
240 ;
241 for (int i = 0; i < Num_OpClasses; ++i) {
242 queueResDist.subname(i, opClassStrings[i]);
243 }
244 */
245 numIssuedDist
246 .init(0,totalWidth,1)
247 .name(name() + ".ISSUE:issued_per_cycle")
248 .desc("Number of insts issued each cycle")
249 .flags(pdf)
250 ;
251 /*
252 dist_unissued
253 .init(Num_OpClasses+2)
254 .name(name() + ".ISSUE:unissued_cause")
255 .desc("Reason ready instruction not issued")
256 .flags(pdf | dist)
257 ;
258 for (int i=0; i < (Num_OpClasses + 2); ++i) {
259 dist_unissued.subname(i, unissued_names[i]);
260 }
261 */
262 statIssuedInstType
263 .init(numThreads,Enums::Num_OpClass)
264 .name(name() + ".ISSUE:FU_type")
265 .desc("Type of FU issued")
266 .flags(total | pdf | dist)
267 ;
268 statIssuedInstType.ysubnames(Enums::OpClassStrings);
269
270 //
271 // How long did instructions for a particular FU type wait prior to issue
272 //
273 /*
274 issueDelayDist
275 .init(Num_OpClasses,0,99,2)
276 .name(name() + ".ISSUE:")
277 .desc("cycles from operands ready to issue")
278 .flags(pdf | cdf)
279 ;
280
281 for (int i=0; i<Num_OpClasses; ++i) {
282 std::stringstream subname;
283 subname << opClassStrings[i] << "_delay";
284 issueDelayDist.subname(i, subname.str());
285 }
286 */
287 issueRate
288 .name(name() + ".ISSUE:rate")
289 .desc("Inst issue rate")
290 .flags(total)
291 ;
292 issueRate = iqInstsIssued / cpu->numCycles;
293
294 statFuBusy
295 .init(Num_OpClasses)
296 .name(name() + ".ISSUE:fu_full")
297 .desc("attempts to use FU when none available")
298 .flags(pdf | dist)
299 ;
300 for (int i=0; i < Num_OpClasses; ++i) {
301 statFuBusy.subname(i, Enums::OpClassStrings[i]);
302 }
303
304 fuBusy
305 .init(numThreads)
306 .name(name() + ".ISSUE:fu_busy_cnt")
307 .desc("FU busy when requested")
308 .flags(total)
309 ;
310
311 fuBusyRate
312 .name(name() + ".ISSUE:fu_busy_rate")
313 .desc("FU busy rate (busy events/executed inst)")
314 .flags(total)
315 ;
316 fuBusyRate = fuBusy / iqInstsIssued;
317
318 for ( int i=0; i < numThreads; i++) {
319 // Tell mem dependence unit to reg stats as well.
320 memDepUnit[i].regStats();
321 }
322 }
323
324 template <class Impl>
325 void
326 InstructionQueue<Impl>::resetState()
327 {
328 //Initialize thread IQ counts
329 for (int i = 0; i <numThreads; i++) {
330 count[i] = 0;
331 instList[i].clear();
332 }
333
334 // Initialize the number of free IQ entries.
335 freeEntries = numEntries;
336
337 // Note that in actuality, the registers corresponding to the logical
338 // registers start off as ready. However this doesn't matter for the
339 // IQ as the instruction should have been correctly told if those
340 // registers are ready in rename. Thus it can all be initialized as
341 // unready.
342 for (int i = 0; i < numPhysRegs; ++i) {
343 regScoreboard[i] = false;
344 }
345
346 for (int i = 0; i < numThreads; ++i) {
347 squashedSeqNum[i] = 0;
348 }
349
350 for (int i = 0; i < Num_OpClasses; ++i) {
351 while (!readyInsts[i].empty())
352 readyInsts[i].pop();
353 queueOnList[i] = false;
354 readyIt[i] = listOrder.end();
355 }
356 nonSpecInsts.clear();
357 listOrder.clear();
358 }
359
360 template <class Impl>
361 void
362 InstructionQueue<Impl>::setActiveThreads(std::list<unsigned> *at_ptr)
363 {
364 activeThreads = at_ptr;
365 }
366
367 template <class Impl>
368 void
369 InstructionQueue<Impl>::setIssueToExecuteQueue(TimeBuffer<IssueStruct> *i2e_ptr)
370 {
371 issueToExecuteQueue = i2e_ptr;
372 }
373
374 template <class Impl>
375 void
376 InstructionQueue<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
377 {
378 timeBuffer = tb_ptr;
379
380 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
381 }
382
383 template <class Impl>
384 void
385 InstructionQueue<Impl>::switchOut()
386 {
387 /*
388 if (!instList[0].empty() || (numEntries != freeEntries) ||
389 !readyInsts[0].empty() || !nonSpecInsts.empty() || !listOrder.empty()) {
390 dumpInsts();
391 // assert(0);
392 }
393 */
394 resetState();
395 dependGraph.reset();
396 instsToExecute.clear();
397 switchedOut = true;
398 for (int i = 0; i < numThreads; ++i) {
399 memDepUnit[i].switchOut();
400 }
401 }
402
403 template <class Impl>
404 void
405 InstructionQueue<Impl>::takeOverFrom()
406 {
407 switchedOut = false;
408 }
409
410 template <class Impl>
411 int
412 InstructionQueue<Impl>::entryAmount(int num_threads)
413 {
414 if (iqPolicy == Partitioned) {
415 return numEntries / num_threads;
416 } else {
417 return 0;
418 }
419 }
420
421
422 template <class Impl>
423 void
424 InstructionQueue<Impl>::resetEntries()
425 {
426 if (iqPolicy != Dynamic || numThreads > 1) {
427 int active_threads = activeThreads->size();
428
429 std::list<unsigned>::iterator threads = activeThreads->begin();
430 std::list<unsigned>::iterator end = activeThreads->end();
431
432 while (threads != end) {
433 unsigned tid = *threads++;
434
435 if (iqPolicy == Partitioned) {
436 maxEntries[tid] = numEntries / active_threads;
437 } else if(iqPolicy == Threshold && active_threads == 1) {
438 maxEntries[tid] = numEntries;
439 }
440 }
441 }
442 }
443
444 template <class Impl>
445 unsigned
446 InstructionQueue<Impl>::numFreeEntries()
447 {
448 return freeEntries;
449 }
450
451 template <class Impl>
452 unsigned
453 InstructionQueue<Impl>::numFreeEntries(unsigned tid)
454 {
455 return maxEntries[tid] - count[tid];
456 }
457
458 // Might want to do something more complex if it knows how many instructions
459 // will be issued this cycle.
460 template <class Impl>
461 bool
462 InstructionQueue<Impl>::isFull()
463 {
464 if (freeEntries == 0) {
465 return(true);
466 } else {
467 return(false);
468 }
469 }
470
471 template <class Impl>
472 bool
473 InstructionQueue<Impl>::isFull(unsigned tid)
474 {
475 if (numFreeEntries(tid) == 0) {
476 return(true);
477 } else {
478 return(false);
479 }
480 }
481
482 template <class Impl>
483 bool
484 InstructionQueue<Impl>::hasReadyInsts()
485 {
486 if (!listOrder.empty()) {
487 return true;
488 }
489
490 for (int i = 0; i < Num_OpClasses; ++i) {
491 if (!readyInsts[i].empty()) {
492 return true;
493 }
494 }
495
496 return false;
497 }
498
499 template <class Impl>
500 void
501 InstructionQueue<Impl>::insert(DynInstPtr &new_inst)
502 {
503 // Make sure the instruction is valid
504 assert(new_inst);
505
506 DPRINTF(IQ, "Adding instruction [sn:%lli] PC %#x to the IQ.\n",
507 new_inst->seqNum, new_inst->readPC());
508
509 assert(freeEntries != 0);
510
511 instList[new_inst->threadNumber].push_back(new_inst);
512
513 --freeEntries;
514
515 new_inst->setInIQ();
516
517 // Look through its source registers (physical regs), and mark any
518 // dependencies.
519 addToDependents(new_inst);
520
521 // Have this instruction set itself as the producer of its destination
522 // register(s).
523 addToProducers(new_inst);
524
525 if (new_inst->isMemRef()) {
526 memDepUnit[new_inst->threadNumber].insert(new_inst);
527 } else {
528 addIfReady(new_inst);
529 }
530
531 ++iqInstsAdded;
532
533 count[new_inst->threadNumber]++;
534
535 assert(freeEntries == (numEntries - countInsts()));
536 }
537
538 template <class Impl>
539 void
540 InstructionQueue<Impl>::insertNonSpec(DynInstPtr &new_inst)
541 {
542 // @todo: Clean up this code; can do it by setting inst as unable
543 // to issue, then calling normal insert on the inst.
544
545 assert(new_inst);
546
547 nonSpecInsts[new_inst->seqNum] = new_inst;
548
549 DPRINTF(IQ, "Adding non-speculative instruction [sn:%lli] PC %#x "
550 "to the IQ.\n",
551 new_inst->seqNum, new_inst->readPC());
552
553 assert(freeEntries != 0);
554
555 instList[new_inst->threadNumber].push_back(new_inst);
556
557 --freeEntries;
558
559 new_inst->setInIQ();
560
561 // Have this instruction set itself as the producer of its destination
562 // register(s).
563 addToProducers(new_inst);
564
565 // If it's a memory instruction, add it to the memory dependency
566 // unit.
567 if (new_inst->isMemRef()) {
568 memDepUnit[new_inst->threadNumber].insertNonSpec(new_inst);
569 }
570
571 ++iqNonSpecInstsAdded;
572
573 count[new_inst->threadNumber]++;
574
575 assert(freeEntries == (numEntries - countInsts()));
576 }
577
578 template <class Impl>
579 void
580 InstructionQueue<Impl>::insertBarrier(DynInstPtr &barr_inst)
581 {
582 memDepUnit[barr_inst->threadNumber].insertBarrier(barr_inst);
583
584 insertNonSpec(barr_inst);
585 }
586
587 template <class Impl>
588 typename Impl::DynInstPtr
589 InstructionQueue<Impl>::getInstToExecute()
590 {
591 assert(!instsToExecute.empty());
592 DynInstPtr inst = instsToExecute.front();
593 instsToExecute.pop_front();
594 return inst;
595 }
596
597 template <class Impl>
598 void
599 InstructionQueue<Impl>::addToOrderList(OpClass op_class)
600 {
601 assert(!readyInsts[op_class].empty());
602
603 ListOrderEntry queue_entry;
604
605 queue_entry.queueType = op_class;
606
607 queue_entry.oldestInst = readyInsts[op_class].top()->seqNum;
608
609 ListOrderIt list_it = listOrder.begin();
610 ListOrderIt list_end_it = listOrder.end();
611
612 while (list_it != list_end_it) {
613 if ((*list_it).oldestInst > queue_entry.oldestInst) {
614 break;
615 }
616
617 list_it++;
618 }
619
620 readyIt[op_class] = listOrder.insert(list_it, queue_entry);
621 queueOnList[op_class] = true;
622 }
623
624 template <class Impl>
625 void
626 InstructionQueue<Impl>::moveToYoungerInst(ListOrderIt list_order_it)
627 {
628 // Get iterator of next item on the list
629 // Delete the original iterator
630 // Determine if the next item is either the end of the list or younger
631 // than the new instruction. If so, then add in a new iterator right here.
632 // If not, then move along.
633 ListOrderEntry queue_entry;
634 OpClass op_class = (*list_order_it).queueType;
635 ListOrderIt next_it = list_order_it;
636
637 ++next_it;
638
639 queue_entry.queueType = op_class;
640 queue_entry.oldestInst = readyInsts[op_class].top()->seqNum;
641
642 while (next_it != listOrder.end() &&
643 (*next_it).oldestInst < queue_entry.oldestInst) {
644 ++next_it;
645 }
646
647 readyIt[op_class] = listOrder.insert(next_it, queue_entry);
648 }
649
650 template <class Impl>
651 void
652 InstructionQueue<Impl>::processFUCompletion(DynInstPtr &inst, int fu_idx)
653 {
654 DPRINTF(IQ, "Processing FU completion [sn:%lli]\n", inst->seqNum);
655 // The CPU could have been sleeping until this op completed (*extremely*
656 // long latency op). Wake it if it was. This may be overkill.
657 if (isSwitchedOut()) {
658 DPRINTF(IQ, "FU completion not processed, IQ is switched out [sn:%lli]\n",
659 inst->seqNum);
660 return;
661 }
662
663 iewStage->wakeCPU();
664
665 if (fu_idx > -1)
666 fuPool->freeUnitNextCycle(fu_idx);
667
668 // @todo: Ensure that these FU Completions happen at the beginning
669 // of a cycle, otherwise they could add too many instructions to
670 // the queue.
671 issueToExecuteQueue->access(-1)->size++;
672 instsToExecute.push_back(inst);
673 }
674
675 // @todo: Figure out a better way to remove the squashed items from the
676 // lists. Checking the top item of each list to see if it's squashed
677 // wastes time and forces jumps.
678 template <class Impl>
679 void
680 InstructionQueue<Impl>::scheduleReadyInsts()
681 {
682 DPRINTF(IQ, "Attempting to schedule ready instructions from "
683 "the IQ.\n");
684
685 IssueStruct *i2e_info = issueToExecuteQueue->access(0);
686
687 // Have iterator to head of the list
688 // While I haven't exceeded bandwidth or reached the end of the list,
689 // Try to get a FU that can do what this op needs.
690 // If successful, change the oldestInst to the new top of the list, put
691 // the queue in the proper place in the list.
692 // Increment the iterator.
693 // This will avoid trying to schedule a certain op class if there are no
694 // FUs that handle it.
695 ListOrderIt order_it = listOrder.begin();
696 ListOrderIt order_end_it = listOrder.end();
697 int total_issued = 0;
698
699 while (total_issued < totalWidth &&
700 iewStage->canIssue() &&
701 order_it != order_end_it) {
702 OpClass op_class = (*order_it).queueType;
703
704 assert(!readyInsts[op_class].empty());
705
706 DynInstPtr issuing_inst = readyInsts[op_class].top();
707
708 assert(issuing_inst->seqNum == (*order_it).oldestInst);
709
710 if (issuing_inst->isSquashed()) {
711 readyInsts[op_class].pop();
712
713 if (!readyInsts[op_class].empty()) {
714 moveToYoungerInst(order_it);
715 } else {
716 readyIt[op_class] = listOrder.end();
717 queueOnList[op_class] = false;
718 }
719
720 listOrder.erase(order_it++);
721
722 ++iqSquashedInstsIssued;
723
724 continue;
725 }
726
727 int idx = -2;
728 int op_latency = 1;
729 int tid = issuing_inst->threadNumber;
730
731 if (op_class != No_OpClass) {
732 idx = fuPool->getUnit(op_class);
733
734 if (idx > -1) {
735 op_latency = fuPool->getOpLatency(op_class);
736 }
737 }
738
739 // If we have an instruction that doesn't require a FU, or a
740 // valid FU, then schedule for execution.
741 if (idx == -2 || idx != -1) {
742 if (op_latency == 1) {
743 i2e_info->size++;
744 instsToExecute.push_back(issuing_inst);
745
746 // Add the FU onto the list of FU's to be freed next
747 // cycle if we used one.
748 if (idx >= 0)
749 fuPool->freeUnitNextCycle(idx);
750 } else {
751 int issue_latency = fuPool->getIssueLatency(op_class);
752 // Generate completion event for the FU
753 FUCompletion *execution = new FUCompletion(issuing_inst,
754 idx, this);
755
756 cpu->schedule(execution, curTick + cpu->ticks(op_latency - 1));
757
758 // @todo: Enforce that issue_latency == 1 or op_latency
759 if (issue_latency > 1) {
760 // If FU isn't pipelined, then it must be freed
761 // upon the execution completing.
762 execution->setFreeFU();
763 } else {
764 // Add the FU onto the list of FU's to be freed next cycle.
765 fuPool->freeUnitNextCycle(idx);
766 }
767 }
768
769 DPRINTF(IQ, "Thread %i: Issuing instruction PC %#x "
770 "[sn:%lli]\n",
771 tid, issuing_inst->readPC(),
772 issuing_inst->seqNum);
773
774 readyInsts[op_class].pop();
775
776 if (!readyInsts[op_class].empty()) {
777 moveToYoungerInst(order_it);
778 } else {
779 readyIt[op_class] = listOrder.end();
780 queueOnList[op_class] = false;
781 }
782
783 issuing_inst->setIssued();
784 ++total_issued;
785
786 if (!issuing_inst->isMemRef()) {
787 // Memory instructions can not be freed from the IQ until they
788 // complete.
789 ++freeEntries;
790 count[tid]--;
791 issuing_inst->clearInIQ();
792 } else {
793 memDepUnit[tid].issue(issuing_inst);
794 }
795
796 listOrder.erase(order_it++);
797 statIssuedInstType[tid][op_class]++;
798 iewStage->incrWb(issuing_inst->seqNum);
799 } else {
800 statFuBusy[op_class]++;
801 fuBusy[tid]++;
802 ++order_it;
803 }
804 }
805
806 numIssuedDist.sample(total_issued);
807 iqInstsIssued+= total_issued;
808
809 // If we issued any instructions, tell the CPU we had activity.
810 if (total_issued) {
811 cpu->activityThisCycle();
812 } else {
813 DPRINTF(IQ, "Not able to schedule any instructions.\n");
814 }
815 }
816
817 template <class Impl>
818 void
819 InstructionQueue<Impl>::scheduleNonSpec(const InstSeqNum &inst)
820 {
821 DPRINTF(IQ, "Marking nonspeculative instruction [sn:%lli] as ready "
822 "to execute.\n", inst);
823
824 NonSpecMapIt inst_it = nonSpecInsts.find(inst);
825
826 assert(inst_it != nonSpecInsts.end());
827
828 unsigned tid = (*inst_it).second->threadNumber;
829
830 (*inst_it).second->setAtCommit();
831
832 (*inst_it).second->setCanIssue();
833
834 if (!(*inst_it).second->isMemRef()) {
835 addIfReady((*inst_it).second);
836 } else {
837 memDepUnit[tid].nonSpecInstReady((*inst_it).second);
838 }
839
840 (*inst_it).second = NULL;
841
842 nonSpecInsts.erase(inst_it);
843 }
844
845 template <class Impl>
846 void
847 InstructionQueue<Impl>::commit(const InstSeqNum &inst, unsigned tid)
848 {
849 DPRINTF(IQ, "[tid:%i]: Committing instructions older than [sn:%i]\n",
850 tid,inst);
851
852 ListIt iq_it = instList[tid].begin();
853
854 while (iq_it != instList[tid].end() &&
855 (*iq_it)->seqNum <= inst) {
856 ++iq_it;
857 instList[tid].pop_front();
858 }
859
860 assert(freeEntries == (numEntries - countInsts()));
861 }
862
863 template <class Impl>
864 int
865 InstructionQueue<Impl>::wakeDependents(DynInstPtr &completed_inst)
866 {
867 int dependents = 0;
868
869 DPRINTF(IQ, "Waking dependents of completed instruction.\n");
870
871 assert(!completed_inst->isSquashed());
872
873 // Tell the memory dependence unit to wake any dependents on this
874 // instruction if it is a memory instruction. Also complete the memory
875 // instruction at this point since we know it executed without issues.
876 // @todo: Might want to rename "completeMemInst" to something that
877 // indicates that it won't need to be replayed, and call this
878 // earlier. Might not be a big deal.
879 if (completed_inst->isMemRef()) {
880 memDepUnit[completed_inst->threadNumber].wakeDependents(completed_inst);
881 completeMemInst(completed_inst);
882 } else if (completed_inst->isMemBarrier() ||
883 completed_inst->isWriteBarrier()) {
884 memDepUnit[completed_inst->threadNumber].completeBarrier(completed_inst);
885 }
886
887 for (int dest_reg_idx = 0;
888 dest_reg_idx < completed_inst->numDestRegs();
889 dest_reg_idx++)
890 {
891 PhysRegIndex dest_reg =
892 completed_inst->renamedDestRegIdx(dest_reg_idx);
893
894 // Special case of uniq or control registers. They are not
895 // handled by the IQ and thus have no dependency graph entry.
896 // @todo Figure out a cleaner way to handle this.
897 if (dest_reg >= numPhysRegs) {
898 continue;
899 }
900
901 DPRINTF(IQ, "Waking any dependents on register %i.\n",
902 (int) dest_reg);
903
904 //Go through the dependency chain, marking the registers as
905 //ready within the waiting instructions.
906 DynInstPtr dep_inst = dependGraph.pop(dest_reg);
907
908 while (dep_inst) {
909 DPRINTF(IQ, "Waking up a dependent instruction, PC%#x.\n",
910 dep_inst->readPC());
911
912 // Might want to give more information to the instruction
913 // so that it knows which of its source registers is
914 // ready. However that would mean that the dependency
915 // graph entries would need to hold the src_reg_idx.
916 dep_inst->markSrcRegReady();
917
918 addIfReady(dep_inst);
919
920 dep_inst = dependGraph.pop(dest_reg);
921
922 ++dependents;
923 }
924
925 // Reset the head node now that all of its dependents have
926 // been woken up.
927 assert(dependGraph.empty(dest_reg));
928 dependGraph.clearInst(dest_reg);
929
930 // Mark the scoreboard as having that register ready.
931 regScoreboard[dest_reg] = true;
932 }
933 return dependents;
934 }
935
936 template <class Impl>
937 void
938 InstructionQueue<Impl>::addReadyMemInst(DynInstPtr &ready_inst)
939 {
940 OpClass op_class = ready_inst->opClass();
941
942 readyInsts[op_class].push(ready_inst);
943
944 // Will need to reorder the list if either a queue is not on the list,
945 // or it has an older instruction than last time.
946 if (!queueOnList[op_class]) {
947 addToOrderList(op_class);
948 } else if (readyInsts[op_class].top()->seqNum <
949 (*readyIt[op_class]).oldestInst) {
950 listOrder.erase(readyIt[op_class]);
951 addToOrderList(op_class);
952 }
953
954 DPRINTF(IQ, "Instruction is ready to issue, putting it onto "
955 "the ready list, PC %#x opclass:%i [sn:%lli].\n",
956 ready_inst->readPC(), op_class, ready_inst->seqNum);
957 }
958
959 template <class Impl>
960 void
961 InstructionQueue<Impl>::rescheduleMemInst(DynInstPtr &resched_inst)
962 {
963 DPRINTF(IQ, "Rescheduling mem inst [sn:%lli]\n", resched_inst->seqNum);
964 resched_inst->clearCanIssue();
965 memDepUnit[resched_inst->threadNumber].reschedule(resched_inst);
966 }
967
968 template <class Impl>
969 void
970 InstructionQueue<Impl>::replayMemInst(DynInstPtr &replay_inst)
971 {
972 memDepUnit[replay_inst->threadNumber].replay(replay_inst);
973 }
974
975 template <class Impl>
976 void
977 InstructionQueue<Impl>::completeMemInst(DynInstPtr &completed_inst)
978 {
979 int tid = completed_inst->threadNumber;
980
981 DPRINTF(IQ, "Completing mem instruction PC:%#x [sn:%lli]\n",
982 completed_inst->readPC(), completed_inst->seqNum);
983
984 ++freeEntries;
985
986 completed_inst->memOpDone = true;
987
988 memDepUnit[tid].completed(completed_inst);
989 count[tid]--;
990 }
991
992 template <class Impl>
993 void
994 InstructionQueue<Impl>::violation(DynInstPtr &store,
995 DynInstPtr &faulting_load)
996 {
997 memDepUnit[store->threadNumber].violation(store, faulting_load);
998 }
999
1000 template <class Impl>
1001 void
1002 InstructionQueue<Impl>::squash(unsigned tid)
1003 {
1004 DPRINTF(IQ, "[tid:%i]: Starting to squash instructions in "
1005 "the IQ.\n", tid);
1006
1007 // Read instruction sequence number of last instruction out of the
1008 // time buffer.
1009 squashedSeqNum[tid] = fromCommit->commitInfo[tid].doneSeqNum;
1010
1011 // Call doSquash if there are insts in the IQ
1012 if (count[tid] > 0) {
1013 doSquash(tid);
1014 }
1015
1016 // Also tell the memory dependence unit to squash.
1017 memDepUnit[tid].squash(squashedSeqNum[tid], tid);
1018 }
1019
1020 template <class Impl>
1021 void
1022 InstructionQueue<Impl>::doSquash(unsigned tid)
1023 {
1024 // Start at the tail.
1025 ListIt squash_it = instList[tid].end();
1026 --squash_it;
1027
1028 DPRINTF(IQ, "[tid:%i]: Squashing until sequence number %i!\n",
1029 tid, squashedSeqNum[tid]);
1030
1031 // Squash any instructions younger than the squashed sequence number
1032 // given.
1033 while (squash_it != instList[tid].end() &&
1034 (*squash_it)->seqNum > squashedSeqNum[tid]) {
1035
1036 DynInstPtr squashed_inst = (*squash_it);
1037
1038 // Only handle the instruction if it actually is in the IQ and
1039 // hasn't already been squashed in the IQ.
1040 if (squashed_inst->threadNumber != tid ||
1041 squashed_inst->isSquashedInIQ()) {
1042 --squash_it;
1043 continue;
1044 }
1045
1046 if (!squashed_inst->isIssued() ||
1047 (squashed_inst->isMemRef() &&
1048 !squashed_inst->memOpDone)) {
1049
1050 DPRINTF(IQ, "[tid:%i]: Instruction [sn:%lli] PC %#x "
1051 "squashed.\n",
1052 tid, squashed_inst->seqNum, squashed_inst->readPC());
1053
1054 // Remove the instruction from the dependency list.
1055 if (!squashed_inst->isNonSpeculative() &&
1056 !squashed_inst->isStoreConditional() &&
1057 !squashed_inst->isMemBarrier() &&
1058 !squashed_inst->isWriteBarrier()) {
1059
1060 for (int src_reg_idx = 0;
1061 src_reg_idx < squashed_inst->numSrcRegs();
1062 src_reg_idx++)
1063 {
1064 PhysRegIndex src_reg =
1065 squashed_inst->renamedSrcRegIdx(src_reg_idx);
1066
1067 // Only remove it from the dependency graph if it
1068 // was placed there in the first place.
1069
1070 // Instead of doing a linked list traversal, we
1071 // can just remove these squashed instructions
1072 // either at issue time, or when the register is
1073 // overwritten. The only downside to this is it
1074 // leaves more room for error.
1075
1076 if (!squashed_inst->isReadySrcRegIdx(src_reg_idx) &&
1077 src_reg < numPhysRegs) {
1078 dependGraph.remove(src_reg, squashed_inst);
1079 }
1080
1081
1082 ++iqSquashedOperandsExamined;
1083 }
1084 } else if (!squashed_inst->isStoreConditional() ||
1085 !squashed_inst->isCompleted()) {
1086 NonSpecMapIt ns_inst_it =
1087 nonSpecInsts.find(squashed_inst->seqNum);
1088 assert(ns_inst_it != nonSpecInsts.end());
1089 if (ns_inst_it == nonSpecInsts.end()) {
1090 assert(squashed_inst->getFault() != NoFault);
1091 } else {
1092
1093 (*ns_inst_it).second = NULL;
1094
1095 nonSpecInsts.erase(ns_inst_it);
1096
1097 ++iqSquashedNonSpecRemoved;
1098 }
1099 }
1100
1101 // Might want to also clear out the head of the dependency graph.
1102
1103 // Mark it as squashed within the IQ.
1104 squashed_inst->setSquashedInIQ();
1105
1106 // @todo: Remove this hack where several statuses are set so the
1107 // inst will flow through the rest of the pipeline.
1108 squashed_inst->setIssued();
1109 squashed_inst->setCanCommit();
1110 squashed_inst->clearInIQ();
1111
1112 //Update Thread IQ Count
1113 count[squashed_inst->threadNumber]--;
1114
1115 ++freeEntries;
1116 }
1117
1118 instList[tid].erase(squash_it--);
1119 ++iqSquashedInstsExamined;
1120 }
1121 }
1122
1123 template <class Impl>
1124 bool
1125 InstructionQueue<Impl>::addToDependents(DynInstPtr &new_inst)
1126 {
1127 // Loop through the instruction's source registers, adding
1128 // them to the dependency list if they are not ready.
1129 int8_t total_src_regs = new_inst->numSrcRegs();
1130 bool return_val = false;
1131
1132 for (int src_reg_idx = 0;
1133 src_reg_idx < total_src_regs;
1134 src_reg_idx++)
1135 {
1136 // Only add it to the dependency graph if it's not ready.
1137 if (!new_inst->isReadySrcRegIdx(src_reg_idx)) {
1138 PhysRegIndex src_reg = new_inst->renamedSrcRegIdx(src_reg_idx);
1139
1140 // Check the IQ's scoreboard to make sure the register
1141 // hasn't become ready while the instruction was in flight
1142 // between stages. Only if it really isn't ready should
1143 // it be added to the dependency graph.
1144 if (src_reg >= numPhysRegs) {
1145 continue;
1146 } else if (regScoreboard[src_reg] == false) {
1147 DPRINTF(IQ, "Instruction PC %#x has src reg %i that "
1148 "is being added to the dependency chain.\n",
1149 new_inst->readPC(), src_reg);
1150
1151 dependGraph.insert(src_reg, new_inst);
1152
1153 // Change the return value to indicate that something
1154 // was added to the dependency graph.
1155 return_val = true;
1156 } else {
1157 DPRINTF(IQ, "Instruction PC %#x has src reg %i that "
1158 "became ready before it reached the IQ.\n",
1159 new_inst->readPC(), src_reg);
1160 // Mark a register ready within the instruction.
1161 new_inst->markSrcRegReady(src_reg_idx);
1162 }
1163 }
1164 }
1165
1166 return return_val;
1167 }
1168
1169 template <class Impl>
1170 void
1171 InstructionQueue<Impl>::addToProducers(DynInstPtr &new_inst)
1172 {
1173 // Nothing really needs to be marked when an instruction becomes
1174 // the producer of a register's value, but for convenience a ptr
1175 // to the producing instruction will be placed in the head node of
1176 // the dependency links.
1177 int8_t total_dest_regs = new_inst->numDestRegs();
1178
1179 for (int dest_reg_idx = 0;
1180 dest_reg_idx < total_dest_regs;
1181 dest_reg_idx++)
1182 {
1183 PhysRegIndex dest_reg = new_inst->renamedDestRegIdx(dest_reg_idx);
1184
1185 // Instructions that use the misc regs will have a reg number
1186 // higher than the normal physical registers. In this case these
1187 // registers are not renamed, and there is no need to track
1188 // dependencies as these instructions must be executed at commit.
1189 if (dest_reg >= numPhysRegs) {
1190 continue;
1191 }
1192
1193 if (!dependGraph.empty(dest_reg)) {
1194 dependGraph.dump();
1195 panic("Dependency graph %i not empty!", dest_reg);
1196 }
1197
1198 dependGraph.setInst(dest_reg, new_inst);
1199
1200 // Mark the scoreboard to say it's not yet ready.
1201 regScoreboard[dest_reg] = false;
1202 }
1203 }
1204
1205 template <class Impl>
1206 void
1207 InstructionQueue<Impl>::addIfReady(DynInstPtr &inst)
1208 {
1209 // If the instruction now has all of its source registers
1210 // available, then add it to the list of ready instructions.
1211 if (inst->readyToIssue()) {
1212
1213 //Add the instruction to the proper ready list.
1214 if (inst->isMemRef()) {
1215
1216 DPRINTF(IQ, "Checking if memory instruction can issue.\n");
1217
1218 // Message to the mem dependence unit that this instruction has
1219 // its registers ready.
1220 memDepUnit[inst->threadNumber].regsReady(inst);
1221
1222 return;
1223 }
1224
1225 OpClass op_class = inst->opClass();
1226
1227 DPRINTF(IQ, "Instruction is ready to issue, putting it onto "
1228 "the ready list, PC %#x opclass:%i [sn:%lli].\n",
1229 inst->readPC(), op_class, inst->seqNum);
1230
1231 readyInsts[op_class].push(inst);
1232
1233 // Will need to reorder the list if either a queue is not on the list,
1234 // or it has an older instruction than last time.
1235 if (!queueOnList[op_class]) {
1236 addToOrderList(op_class);
1237 } else if (readyInsts[op_class].top()->seqNum <
1238 (*readyIt[op_class]).oldestInst) {
1239 listOrder.erase(readyIt[op_class]);
1240 addToOrderList(op_class);
1241 }
1242 }
1243 }
1244
1245 template <class Impl>
1246 int
1247 InstructionQueue<Impl>::countInsts()
1248 {
1249 #if 0
1250 //ksewell:This works but definitely could use a cleaner write
1251 //with a more intuitive way of counting. Right now it's
1252 //just brute force ....
1253 // Change the #if if you want to use this method.
1254 int total_insts = 0;
1255
1256 for (int i = 0; i < numThreads; ++i) {
1257 ListIt count_it = instList[i].begin();
1258
1259 while (count_it != instList[i].end()) {
1260 if (!(*count_it)->isSquashed() && !(*count_it)->isSquashedInIQ()) {
1261 if (!(*count_it)->isIssued()) {
1262 ++total_insts;
1263 } else if ((*count_it)->isMemRef() &&
1264 !(*count_it)->memOpDone) {
1265 // Loads that have not been marked as executed still count
1266 // towards the total instructions.
1267 ++total_insts;
1268 }
1269 }
1270
1271 ++count_it;
1272 }
1273 }
1274
1275 return total_insts;
1276 #else
1277 return numEntries - freeEntries;
1278 #endif
1279 }
1280
1281 template <class Impl>
1282 void
1283 InstructionQueue<Impl>::dumpLists()
1284 {
1285 for (int i = 0; i < Num_OpClasses; ++i) {
1286 cprintf("Ready list %i size: %i\n", i, readyInsts[i].size());
1287
1288 cprintf("\n");
1289 }
1290
1291 cprintf("Non speculative list size: %i\n", nonSpecInsts.size());
1292
1293 NonSpecMapIt non_spec_it = nonSpecInsts.begin();
1294 NonSpecMapIt non_spec_end_it = nonSpecInsts.end();
1295
1296 cprintf("Non speculative list: ");
1297
1298 while (non_spec_it != non_spec_end_it) {
1299 cprintf("%#x [sn:%lli]", (*non_spec_it).second->readPC(),
1300 (*non_spec_it).second->seqNum);
1301 ++non_spec_it;
1302 }
1303
1304 cprintf("\n");
1305
1306 ListOrderIt list_order_it = listOrder.begin();
1307 ListOrderIt list_order_end_it = listOrder.end();
1308 int i = 1;
1309
1310 cprintf("List order: ");
1311
1312 while (list_order_it != list_order_end_it) {
1313 cprintf("%i OpClass:%i [sn:%lli] ", i, (*list_order_it).queueType,
1314 (*list_order_it).oldestInst);
1315
1316 ++list_order_it;
1317 ++i;
1318 }
1319
1320 cprintf("\n");
1321 }
1322
1323
1324 template <class Impl>
1325 void
1326 InstructionQueue<Impl>::dumpInsts()
1327 {
1328 for (int i = 0; i < numThreads; ++i) {
1329 int num = 0;
1330 int valid_num = 0;
1331 ListIt inst_list_it = instList[i].begin();
1332
1333 while (inst_list_it != instList[i].end())
1334 {
1335 cprintf("Instruction:%i\n",
1336 num);
1337 if (!(*inst_list_it)->isSquashed()) {
1338 if (!(*inst_list_it)->isIssued()) {
1339 ++valid_num;
1340 cprintf("Count:%i\n", valid_num);
1341 } else if ((*inst_list_it)->isMemRef() &&
1342 !(*inst_list_it)->memOpDone) {
1343 // Loads that have not been marked as executed
1344 // still count towards the total instructions.
1345 ++valid_num;
1346 cprintf("Count:%i\n", valid_num);
1347 }
1348 }
1349
1350 cprintf("PC:%#x\n[sn:%lli]\n[tid:%i]\n"
1351 "Issued:%i\nSquashed:%i\n",
1352 (*inst_list_it)->readPC(),
1353 (*inst_list_it)->seqNum,
1354 (*inst_list_it)->threadNumber,
1355 (*inst_list_it)->isIssued(),
1356 (*inst_list_it)->isSquashed());
1357
1358 if ((*inst_list_it)->isMemRef()) {
1359 cprintf("MemOpDone:%i\n", (*inst_list_it)->memOpDone);
1360 }
1361
1362 cprintf("\n");
1363
1364 inst_list_it++;
1365 ++num;
1366 }
1367 }
1368
1369 cprintf("Insts to Execute list:\n");
1370
1371 int num = 0;
1372 int valid_num = 0;
1373 ListIt inst_list_it = instsToExecute.begin();
1374
1375 while (inst_list_it != instsToExecute.end())
1376 {
1377 cprintf("Instruction:%i\n",
1378 num);
1379 if (!(*inst_list_it)->isSquashed()) {
1380 if (!(*inst_list_it)->isIssued()) {
1381 ++valid_num;
1382 cprintf("Count:%i\n", valid_num);
1383 } else if ((*inst_list_it)->isMemRef() &&
1384 !(*inst_list_it)->memOpDone) {
1385 // Loads that have not been marked as executed
1386 // still count towards the total instructions.
1387 ++valid_num;
1388 cprintf("Count:%i\n", valid_num);
1389 }
1390 }
1391
1392 cprintf("PC:%#x\n[sn:%lli]\n[tid:%i]\n"
1393 "Issued:%i\nSquashed:%i\n",
1394 (*inst_list_it)->readPC(),
1395 (*inst_list_it)->seqNum,
1396 (*inst_list_it)->threadNumber,
1397 (*inst_list_it)->isIssued(),
1398 (*inst_list_it)->isSquashed());
1399
1400 if ((*inst_list_it)->isMemRef()) {
1401 cprintf("MemOpDone:%i\n", (*inst_list_it)->memOpDone);
1402 }
1403
1404 cprintf("\n");
1405
1406 inst_list_it++;
1407 ++num;
1408 }
1409 }