1/*
2 * Copyright (c) 2013-2014 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 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions are
16 * met: redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer;
18 * redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution;
21 * neither the name of the copyright holders nor the names of its
22 * contributors may be used to endorse or promote products derived from
23 * this software without specific prior written permission.
24 *
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
29 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
30 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
31 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
32 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
33 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
34 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
35 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 *
37 * Authors: Andrew Bardsley
38 */
39
40#include "cpu/minor/execute.hh"
41
42#include "arch/locked_mem.hh"
43#include "arch/registers.hh"
44#include "arch/utility.hh"
45#include "cpu/minor/cpu.hh"
46#include "cpu/minor/exec_context.hh"
47#include "cpu/minor/fetch1.hh"
48#include "cpu/minor/lsq.hh"
49#include "cpu/op_class.hh"
50#include "debug/Activity.hh"
51#include "debug/Branch.hh"
52#include "debug/Drain.hh"
53#include "debug/MinorExecute.hh"
54#include "debug/MinorInterrupt.hh"
55#include "debug/MinorMem.hh"
56#include "debug/MinorTrace.hh"
57#include "debug/PCEvent.hh"
58
59namespace Minor
60{
61
62Execute::Execute(const std::string &name_,
63 MinorCPU &cpu_,
64 MinorCPUParams &params,
65 Latch<ForwardInstData>::Output inp_,
66 Latch<BranchData>::Input out_) :
67 Named(name_),
68 inp(inp_),
69 out(out_),
70 cpu(cpu_),
71 issueLimit(params.executeIssueLimit),
72 memoryIssueLimit(params.executeMemoryIssueLimit),
73 commitLimit(params.executeCommitLimit),
74 memoryCommitLimit(params.executeMemoryCommitLimit),
75 processMoreThanOneInput(params.executeCycleInput),
76 fuDescriptions(*params.executeFuncUnits),
77 numFuncUnits(fuDescriptions.funcUnits.size()),
78 setTraceTimeOnCommit(params.executeSetTraceTimeOnCommit),
79 setTraceTimeOnIssue(params.executeSetTraceTimeOnIssue),
80 allowEarlyMemIssue(params.executeAllowEarlyMemoryIssue),
81 noCostFUIndex(fuDescriptions.funcUnits.size() + 1),
82 lsq(name_ + ".lsq", name_ + ".dcache_port",
83 cpu_, *this,
84 params.executeMaxAccessesInMemory,
85 params.executeMemoryWidth,
86 params.executeLSQRequestsQueueSize,
87 params.executeLSQTransfersQueueSize,
88 params.executeLSQStoreBufferSize,
89 params.executeLSQMaxStoreBufferStoresPerCycle),
90 executeInfo(params.numThreads, ExecuteThreadInfo(params.executeCommitLimit)),
91 interruptPriority(0),
92 issuePriority(0),
93 commitPriority(0)
94{
95 if (commitLimit < 1) {
96 fatal("%s: executeCommitLimit must be >= 1 (%d)\n", name_,
97 commitLimit);
98 }
99
100 if (issueLimit < 1) {
101 fatal("%s: executeCommitLimit must be >= 1 (%d)\n", name_,
102 issueLimit);
103 }
104
105 if (memoryIssueLimit < 1) {
106 fatal("%s: executeMemoryIssueLimit must be >= 1 (%d)\n", name_,
107 memoryIssueLimit);
108 }
109
110 if (memoryCommitLimit > commitLimit) {
111 fatal("%s: executeMemoryCommitLimit (%d) must be <="
112 " executeCommitLimit (%d)\n",
113 name_, memoryCommitLimit, commitLimit);
114 }
115
116 if (params.executeInputBufferSize < 1) {
117 fatal("%s: executeInputBufferSize must be >= 1 (%d)\n", name_,
118 params.executeInputBufferSize);
119 }
120
121 if (params.executeInputBufferSize < 1) {
122 fatal("%s: executeInputBufferSize must be >= 1 (%d)\n", name_,
123 params.executeInputBufferSize);
124 }
125
126 /* This should be large enough to count all the in-FU instructions
127 * which need to be accounted for in the inFlightInsts
128 * queue */
129 unsigned int total_slots = 0;
130
131 /* Make FUPipelines for each MinorFU */
132 for (unsigned int i = 0; i < numFuncUnits; i++) {
133 std::ostringstream fu_name;
134 MinorFU *fu_description = fuDescriptions.funcUnits[i];
135
136 /* Note the total number of instruction slots (for sizing
137 * the inFlightInst queue) and the maximum latency of any FU
138 * (for sizing the activity recorder) */
139 total_slots += fu_description->opLat;
140
141 fu_name << name_ << ".fu." << i;
142
143 FUPipeline *fu = new FUPipeline(fu_name.str(), *fu_description, cpu);
144
145 funcUnits.push_back(fu);
146 }
147
148 /** Check that there is a functional unit for all operation classes */
149 for (int op_class = No_OpClass + 1; op_class < Num_OpClasses; op_class++) {
150 bool found_fu = false;
151 unsigned int fu_index = 0;
152
153 while (fu_index < numFuncUnits && !found_fu)
154 {
155 if (funcUnits[fu_index]->provides(
156 static_cast<OpClass>(op_class)))
157 {
158 found_fu = true;
159 }
160 fu_index++;
161 }
162
163 if (!found_fu) {
164 warn("No functional unit for OpClass %s\n",
165 Enums::OpClassStrings[op_class]);
166 }
167 }
168
169 /* Per-thread structures */
170 for (ThreadID tid = 0; tid < params.numThreads; tid++) {
171 std::string tid_str = std::to_string(tid);
172
173 /* Input Buffers */
174 inputBuffer.push_back(
175 InputBuffer<ForwardInstData>(
176 name_ + ".inputBuffer" + tid_str, "insts",
177 params.executeInputBufferSize));
178
179 /* Scoreboards */
180 scoreboard.push_back(Scoreboard(name_ + ".scoreboard" + tid_str));
181
182 /* In-flight instruction records */
183 executeInfo[tid].inFlightInsts = new Queue<QueuedInst,
184 ReportTraitsAdaptor<QueuedInst> >(
185 name_ + ".inFlightInsts" + tid_str, "insts", total_slots);
186
187 executeInfo[tid].inFUMemInsts = new Queue<QueuedInst,
188 ReportTraitsAdaptor<QueuedInst> >(
189 name_ + ".inFUMemInsts" + tid_str, "insts", total_slots);
190 }
191}
192
193const ForwardInstData *
194Execute::getInput(ThreadID tid)
195{
196 /* Get a line from the inputBuffer to work with */
197 if (!inputBuffer[tid].empty()) {
198 const ForwardInstData &head = inputBuffer[tid].front();
199
200 return (head.isBubble() ? NULL : &(inputBuffer[tid].front()));
201 } else {
202 return NULL;
203 }
204}
205
206void
207Execute::popInput(ThreadID tid)
208{
209 if (!inputBuffer[tid].empty())
210 inputBuffer[tid].pop();
211
212 executeInfo[tid].inputIndex = 0;
213}
214
215void
216Execute::tryToBranch(MinorDynInstPtr inst, Fault fault, BranchData &branch)
217{
218 ThreadContext *thread = cpu.getContext(inst->id.threadId);
219 const TheISA::PCState &pc_before = inst->pc;
220 TheISA::PCState target = thread->pcState();
221
222 /* Force a branch for SerializeAfter/SquashAfter instructions
223 * at the end of micro-op sequence when we're not suspended */
224 bool force_branch = thread->status() != ThreadContext::Suspended &&
225 !inst->isFault() &&
226 inst->isLastOpInInst() &&
227 (inst->staticInst->isSerializeAfter() ||
228 inst->staticInst->isSquashAfter() ||
229 inst->staticInst->isIprAccess());
230
231 DPRINTF(Branch, "tryToBranch before: %s after: %s%s\n",
232 pc_before, target, (force_branch ? " (forcing)" : ""));
233
234 /* Will we change the PC to something other than the next instruction? */
235 bool must_branch = pc_before != target ||
236 fault != NoFault ||
237 force_branch;
238
239 /* The reason for the branch data we're about to generate, set below */
240 BranchData::Reason reason = BranchData::NoBranch;
241
242 if (fault == NoFault)
243 {
244 TheISA::advancePC(target, inst->staticInst);
245 thread->pcState(target);
246
247 DPRINTF(Branch, "Advancing current PC from: %s to: %s\n",
248 pc_before, target);
249 }
250
251 if (inst->predictedTaken && !force_branch) {
252 /* Predicted to branch */
253 if (!must_branch) {
254 /* No branch was taken, change stream to get us back to the
255 * intended PC value */
256 DPRINTF(Branch, "Predicted a branch from 0x%x to 0x%x but"
257 " none happened inst: %s\n",
258 inst->pc.instAddr(), inst->predictedTarget.instAddr(), *inst);
259
260 reason = BranchData::BadlyPredictedBranch;
261 } else if (inst->predictedTarget == target) {
262 /* Branch prediction got the right target, kill the branch and
263 * carry on.
264 * Note that this information to the branch predictor might get
265 * overwritten by a "real" branch during this cycle */
266 DPRINTF(Branch, "Predicted a branch from 0x%x to 0x%x correctly"
267 " inst: %s\n",
268 inst->pc.instAddr(), inst->predictedTarget.instAddr(), *inst);
269
270 reason = BranchData::CorrectlyPredictedBranch;
271 } else {
272 /* Branch prediction got the wrong target */
273 DPRINTF(Branch, "Predicted a branch from 0x%x to 0x%x"
274 " but got the wrong target (actual: 0x%x) inst: %s\n",
275 inst->pc.instAddr(), inst->predictedTarget.instAddr(),
276 target.instAddr(), *inst);
277
278 reason = BranchData::BadlyPredictedBranchTarget;
279 }
280 } else if (must_branch) {
281 /* Unpredicted branch */
282 DPRINTF(Branch, "Unpredicted branch from 0x%x to 0x%x inst: %s\n",
283 inst->pc.instAddr(), target.instAddr(), *inst);
284
285 reason = BranchData::UnpredictedBranch;
286 } else {
287 /* No branch at all */
288 reason = BranchData::NoBranch;
289 }
290
291 updateBranchData(inst->id.threadId, reason, inst, target, branch);
292}
293
294void
295Execute::updateBranchData(
296 ThreadID tid,
297 BranchData::Reason reason,
298 MinorDynInstPtr inst, const TheISA::PCState &target,
299 BranchData &branch)
300{
301 if (reason != BranchData::NoBranch) {
302 /* Bump up the stream sequence number on a real branch*/
303 if (BranchData::isStreamChange(reason))
304 executeInfo[tid].streamSeqNum++;
305
306 /* Branches (even mis-predictions) don't change the predictionSeqNum,
307 * just the streamSeqNum */
308 branch = BranchData(reason, tid,
309 executeInfo[tid].streamSeqNum,
310 /* Maintaining predictionSeqNum if there's no inst is just a
311 * courtesy and looks better on minorview */
312 (inst->isBubble() ? executeInfo[tid].lastPredictionSeqNum
313 : inst->id.predictionSeqNum),
314 target, inst);
315
316 DPRINTF(Branch, "Branch data signalled: %s\n", branch);
317 }
318}
319
320void
321Execute::handleMemResponse(MinorDynInstPtr inst,
322 LSQ::LSQRequestPtr response, BranchData &branch, Fault &fault)
323{
324 ThreadID thread_id = inst->id.threadId;
325 ThreadContext *thread = cpu.getContext(thread_id);
326
327 ExecContext context(cpu, *cpu.threads[thread_id], *this, inst);
328
329 PacketPtr packet = response->packet;
330
331 bool is_load = inst->staticInst->isLoad();
332 bool is_store = inst->staticInst->isStore();
333 bool is_prefetch = inst->staticInst->isDataPrefetch();
334
335 /* If true, the trace's predicate value will be taken from the exec
336 * context predicate, otherwise, it will be set to false */
337 bool use_context_predicate = true;
338
339 if (response->fault != NoFault) {
340 /* Invoke memory faults. */
341 DPRINTF(MinorMem, "Completing fault from DTLB access: %s\n",
342 response->fault->name());
343
344 if (inst->staticInst->isPrefetch()) {
345 DPRINTF(MinorMem, "Not taking fault on prefetch: %s\n",
346 response->fault->name());
347
348 /* Don't assign to fault */
349 } else {
350 /* Take the fault raised during the TLB/memory access */
351 fault = response->fault;
352
353 fault->invoke(thread, inst->staticInst);
354 }
355 } else if (!packet) {
356 DPRINTF(MinorMem, "Completing failed request inst: %s\n",
357 *inst);
358 use_context_predicate = false;
359 } else if (packet->isError()) {
360 DPRINTF(MinorMem, "Trying to commit error response: %s\n",
361 *inst);
362
363 fatal("Received error response packet for inst: %s\n", *inst);
364 } else if (is_store || is_load || is_prefetch) {
365 assert(packet);
366
367 DPRINTF(MinorMem, "Memory response inst: %s addr: 0x%x size: %d\n",
368 *inst, packet->getAddr(), packet->getSize());
369
370 if (is_load && packet->getSize() > 0) {
371 DPRINTF(MinorMem, "Memory data[0]: 0x%x\n",
372 static_cast<unsigned int>(packet->getConstPtr<uint8_t>()[0]));
373 }
374
375 /* Complete the memory access instruction */
376 fault = inst->staticInst->completeAcc(packet, &context,
377 inst->traceData);
378
379 if (fault != NoFault) {
380 /* Invoke fault created by instruction completion */
381 DPRINTF(MinorMem, "Fault in memory completeAcc: %s\n",
382 fault->name());
383 fault->invoke(thread, inst->staticInst);
384 } else {
385 /* Stores need to be pushed into the store buffer to finish
386 * them off */
387 if (response->needsToBeSentToStoreBuffer())
388 lsq.sendStoreToStoreBuffer(response);
389 }
390 } else {
391 fatal("There should only ever be reads, "
392 "writes or faults at this point\n");
393 }
394
395 lsq.popResponse(response);
396
397 if (inst->traceData) {
398 inst->traceData->setPredicate((use_context_predicate ?
399 context.readPredicate() : false));
400 }
401
402 doInstCommitAccounting(inst);
403
404 /* Generate output to account for branches */
405 tryToBranch(inst, fault, branch);
406}
407
408bool
409Execute::isInterrupted(ThreadID thread_id) const
410{
411 return cpu.checkInterrupts(cpu.getContext(thread_id));
412}
413
414bool
415Execute::takeInterrupt(ThreadID thread_id, BranchData &branch)
416{
417 DPRINTF(MinorInterrupt, "Considering interrupt status from PC: %s\n",
418 cpu.getContext(thread_id)->pcState());
419
420 Fault interrupt = cpu.getInterruptController(thread_id)->getInterrupt
421 (cpu.getContext(thread_id));
422
423 if (interrupt != NoFault) {
424 /* The interrupt *must* set pcState */
425 cpu.getInterruptController(thread_id)->updateIntrInfo
426 (cpu.getContext(thread_id));
427 interrupt->invoke(cpu.getContext(thread_id));
428
429 assert(!lsq.accessesInFlight());
430
431 DPRINTF(MinorInterrupt, "Invoking interrupt: %s to PC: %s\n",
432 interrupt->name(), cpu.getContext(thread_id)->pcState());
433
434 /* Assume that an interrupt *must* cause a branch. Assert this? */
435
436 updateBranchData(thread_id, BranchData::Interrupt,
437 MinorDynInst::bubble(), cpu.getContext(thread_id)->pcState(),
438 branch);
439 }
440
441 return interrupt != NoFault;
442}
443
444bool
445Execute::executeMemRefInst(MinorDynInstPtr inst, BranchData &branch,
446 bool &passed_predicate, Fault &fault)
447{
448 bool issued = false;
449
450 /* Set to true if the mem op. is issued and sent to the mem system */
451 passed_predicate = false;
452
453 if (!lsq.canRequest()) {
454 /* Not acting on instruction yet as the memory
455 * queues are full */
456 issued = false;
457 } else {
458 ThreadContext *thread = cpu.getContext(inst->id.threadId);
459 TheISA::PCState old_pc = thread->pcState();
460
461 ExecContext context(cpu, *cpu.threads[inst->id.threadId],
462 *this, inst);
463
464 DPRINTF(MinorExecute, "Initiating memRef inst: %s\n", *inst);
465
466 Fault init_fault = inst->staticInst->initiateAcc(&context,
467 inst->traceData);
468
469 if (init_fault != NoFault) {
470 DPRINTF(MinorExecute, "Fault on memory inst: %s"
471 " initiateAcc: %s\n", *inst, init_fault->name());
472 fault = init_fault;
473 } else {
474 /* Only set this if the instruction passed its
475 * predicate */
476 passed_predicate = context.readPredicate();
477
478 /* Set predicate in tracing */
479 if (inst->traceData)
480 inst->traceData->setPredicate(passed_predicate);
481
482 /* If the instruction didn't pass its predicate (and so will not
483 * progress from here) Try to branch to correct and branch
484 * mis-prediction. */
485 if (!passed_predicate) {
486 /* Leave it up to commit to handle the fault */
487 lsq.pushFailedRequest(inst);
488 }
489 }
490
491 /* Restore thread PC */
492 thread->pcState(old_pc);
493 issued = true;
494 }
495
496 return issued;
497}
498
499/** Increment a cyclic buffer index for indices [0, cycle_size-1] */
500inline unsigned int
501cyclicIndexInc(unsigned int index, unsigned int cycle_size)
502{
503 unsigned int ret = index + 1;
504
505 if (ret == cycle_size)
506 ret = 0;
507
508 return ret;
509}
510
511/** Decrement a cyclic buffer index for indices [0, cycle_size-1] */
512inline unsigned int
513cyclicIndexDec(unsigned int index, unsigned int cycle_size)
514{
515 int ret = index - 1;
516
517 if (ret < 0)
518 ret = cycle_size - 1;
519
520 return ret;
521}
522
523unsigned int
524Execute::issue(ThreadID thread_id)
525{
526 const ForwardInstData *insts_in = getInput(thread_id);
527 ExecuteThreadInfo &thread = executeInfo[thread_id];
528
529 /* Early termination if we have no instructions */
530 if (!insts_in)
531 return 0;
532
533 /* Start from the first FU */
534 unsigned int fu_index = 0;
535
536 /* Remains true while instructions are still being issued. If any
537 * instruction fails to issue, this is set to false and we exit issue.
538 * This strictly enforces in-order issue. For other issue behaviours,
539 * a more complicated test in the outer while loop below is needed. */
540 bool issued = true;
541
542 /* Number of insts issues this cycle to check for issueLimit */
543 unsigned num_insts_issued = 0;
544
545 /* Number of memory ops issues this cycle to check for memoryIssueLimit */
546 unsigned num_mem_insts_issued = 0;
547
548 /* Number of instructions discarded this cycle in order to enforce a
549 * discardLimit. @todo, add that parameter? */
550 unsigned num_insts_discarded = 0;
551
552 do {
553 MinorDynInstPtr inst = insts_in->insts[thread.inputIndex];
554 Fault fault = inst->fault;
555 bool discarded = false;
556 bool issued_mem_ref = false;
557
558 if (inst->isBubble()) {
559 /* Skip */
560 issued = true;
561 } else if (cpu.getContext(thread_id)->status() ==
562 ThreadContext::Suspended)
563 {
564 DPRINTF(MinorExecute, "Discarding inst: %s from suspended"
565 " thread\n", *inst);
566
567 issued = true;
568 discarded = true;
569 } else if (inst->id.streamSeqNum != thread.streamSeqNum) {
570 DPRINTF(MinorExecute, "Discarding inst: %s as its stream"
571 " state was unexpected, expected: %d\n",
572 *inst, thread.streamSeqNum);
573 issued = true;
574 discarded = true;
575 } else {
576 /* Try and issue an instruction into an FU, assume we didn't and
577 * fix that in the loop */
578 issued = false;
579
580 /* Try FU from 0 each instruction */
581 fu_index = 0;
582
583 /* Try and issue a single instruction stepping through the
584 * available FUs */
585 do {
586 FUPipeline *fu = funcUnits[fu_index];
587
588 DPRINTF(MinorExecute, "Trying to issue inst: %s to FU: %d\n",
589 *inst, fu_index);
590
591 /* Does the examined fu have the OpClass-related capability
592 * needed to execute this instruction? Faults can always
593 * issue to any FU but probably should just 'live' in the
594 * inFlightInsts queue rather than having an FU. */
595 bool fu_is_capable = (!inst->isFault() ?
596 fu->provides(inst->staticInst->opClass()) : true);
597
598 if (inst->isNoCostInst()) {
599 /* Issue free insts. to a fake numbered FU */
600 fu_index = noCostFUIndex;
601
602 /* And start the countdown on activity to allow
603 * this instruction to get to the end of its FU */
604 cpu.activityRecorder->activity();
605
606 /* Mark the destinations for this instruction as
607 * busy */
608 scoreboard[thread_id].markupInstDests(inst, cpu.curCycle() +
609 Cycles(0), cpu.getContext(thread_id), false);
610
611 DPRINTF(MinorExecute, "Issuing %s to %d\n", inst->id, noCostFUIndex);
612 inst->fuIndex = noCostFUIndex;
613 inst->extraCommitDelay = Cycles(0);
614 inst->extraCommitDelayExpr = NULL;
615
616 /* Push the instruction onto the inFlight queue so
617 * it can be committed in order */
618 QueuedInst fu_inst(inst);
619 thread.inFlightInsts->push(fu_inst);
620
621 issued = true;
622
623 } else if (!fu_is_capable || fu->alreadyPushed()) {
624 /* Skip */
625 if (!fu_is_capable) {
626 DPRINTF(MinorExecute, "Can't issue as FU: %d isn't"
627 " capable\n", fu_index);
628 } else {
629 DPRINTF(MinorExecute, "Can't issue as FU: %d is"
630 " already busy\n", fu_index);
631 }
632 } else if (fu->stalled) {
633 DPRINTF(MinorExecute, "Can't issue inst: %s into FU: %d,"
634 " it's stalled\n",
635 *inst, fu_index);
636 } else if (!fu->canInsert()) {
637 DPRINTF(MinorExecute, "Can't issue inst: %s to busy FU"
638 " for another: %d cycles\n",
639 *inst, fu->cyclesBeforeInsert());
640 } else {
641 MinorFUTiming *timing = (!inst->isFault() ?
642 fu->findTiming(inst->staticInst) : NULL);
643
644 const std::vector<Cycles> *src_latencies =
645 (timing ? &(timing->srcRegsRelativeLats)
646 : NULL);
647
648 const std::vector<bool> *cant_forward_from_fu_indices =
649 &(fu->cantForwardFromFUIndices);
650
651 if (timing && timing->suppress) {
652 DPRINTF(MinorExecute, "Can't issue inst: %s as extra"
653 " decoding is suppressing it\n",
654 *inst);
655 } else if (!scoreboard[thread_id].canInstIssue(inst,
656 src_latencies, cant_forward_from_fu_indices,
657 cpu.curCycle(), cpu.getContext(thread_id)))
658 {
659 DPRINTF(MinorExecute, "Can't issue inst: %s yet\n",
660 *inst);
661 } else {
662 /* Can insert the instruction into this FU */
663 DPRINTF(MinorExecute, "Issuing inst: %s"
664 " into FU %d\n", *inst,
665 fu_index);
666
667 Cycles extra_dest_retire_lat = Cycles(0);
668 TimingExpr *extra_dest_retire_lat_expr = NULL;
669 Cycles extra_assumed_lat = Cycles(0);
670
671 /* Add the extraCommitDelay and extraAssumeLat to
672 * the FU pipeline timings */
673 if (timing) {
674 extra_dest_retire_lat =
675 timing->extraCommitLat;
676 extra_dest_retire_lat_expr =
677 timing->extraCommitLatExpr;
678 extra_assumed_lat =
679 timing->extraAssumedLat;
680 }
681
682 issued_mem_ref = inst->isMemRef();
683
684 QueuedInst fu_inst(inst);
685
686 /* Decorate the inst with FU details */
687 inst->fuIndex = fu_index;
688 inst->extraCommitDelay = extra_dest_retire_lat;
689 inst->extraCommitDelayExpr =
690 extra_dest_retire_lat_expr;
691
692 if (issued_mem_ref) {
693 /* Remember which instruction this memory op
694 * depends on so that initiateAcc can be called
695 * early */
696 if (allowEarlyMemIssue) {
697 inst->instToWaitFor =
698 scoreboard[thread_id].execSeqNumToWaitFor(inst,
699 cpu.getContext(thread_id));
700
701 if (lsq.getLastMemBarrier(thread_id) >
702 inst->instToWaitFor)
703 {
704 DPRINTF(MinorExecute, "A barrier will"
705 " cause a delay in mem ref issue of"
706 " inst: %s until after inst"
707 " %d(exec)\n", *inst,
708 lsq.getLastMemBarrier(thread_id));
709
710 inst->instToWaitFor =
711 lsq.getLastMemBarrier(thread_id);
712 } else {
713 DPRINTF(MinorExecute, "Memory ref inst:"
714 " %s must wait for inst %d(exec)"
715 " before issuing\n",
716 *inst, inst->instToWaitFor);
717 }
718
719 inst->canEarlyIssue = true;
720 }
721 /* Also queue this instruction in the memory ref
722 * queue to ensure in-order issue to the LSQ */
723 DPRINTF(MinorExecute, "Pushing mem inst: %s\n",
724 *inst);
725 thread.inFUMemInsts->push(fu_inst);
726 }
727
728 /* Issue to FU */
729 fu->push(fu_inst);
730 /* And start the countdown on activity to allow
731 * this instruction to get to the end of its FU */
732 cpu.activityRecorder->activity();
733
734 /* Mark the destinations for this instruction as
735 * busy */
736 scoreboard[thread_id].markupInstDests(inst, cpu.curCycle() +
737 fu->description.opLat +
738 extra_dest_retire_lat +
739 extra_assumed_lat,
740 cpu.getContext(thread_id),
741 issued_mem_ref && extra_assumed_lat == Cycles(0));
742
743 /* Push the instruction onto the inFlight queue so
744 * it can be committed in order */
745 thread.inFlightInsts->push(fu_inst);
746
747 issued = true;
748 }
749 }
750
751 fu_index++;
752 } while (fu_index != numFuncUnits && !issued);
753
754 if (!issued)
755 DPRINTF(MinorExecute, "Didn't issue inst: %s\n", *inst);
756 }
757
758 if (issued) {
759 /* Generate MinorTrace's MinorInst lines. Do this at commit
760 * to allow better instruction annotation? */
761 if (DTRACE(MinorTrace) && !inst->isBubble())
762 inst->minorTraceInst(*this);
763
764 /* Mark up barriers in the LSQ */
765 if (!discarded && inst->isInst() &&
766 inst->staticInst->isMemBarrier())
767 {
768 DPRINTF(MinorMem, "Issuing memory barrier inst: %s\n", *inst);
769 lsq.issuedMemBarrierInst(inst);
770 }
771
772 if (inst->traceData && setTraceTimeOnIssue) {
773 inst->traceData->setWhen(curTick());
774 }
775
776 if (issued_mem_ref)
777 num_mem_insts_issued++;
778
779 if (discarded) {
780 num_insts_discarded++;
781 } else if (!inst->isBubble()) {
782 num_insts_issued++;
783
784 if (num_insts_issued == issueLimit)
785 DPRINTF(MinorExecute, "Reached inst issue limit\n");
786 }
787
788 thread.inputIndex++;
789 DPRINTF(MinorExecute, "Stepping to next inst inputIndex: %d\n",
790 thread.inputIndex);
791 }
792
793 /* Got to the end of a line */
794 if (thread.inputIndex == insts_in->width()) {
795 popInput(thread_id);
796 /* Set insts_in to null to force us to leave the surrounding
797 * loop */
798 insts_in = NULL;
799
800 if (processMoreThanOneInput) {
801 DPRINTF(MinorExecute, "Wrapping\n");
802 insts_in = getInput(thread_id);
803 }
804 }
805 } while (insts_in && thread.inputIndex < insts_in->width() &&
806 /* We still have instructions */
807 fu_index != numFuncUnits && /* Not visited all FUs */
808 issued && /* We've not yet failed to issue an instruction */
809 num_insts_issued != issueLimit && /* Still allowed to issue */
810 num_mem_insts_issued != memoryIssueLimit);
811
812 return num_insts_issued;
813}
814
815bool
816Execute::tryPCEvents(ThreadID thread_id)
817{
818 ThreadContext *thread = cpu.getContext(thread_id);
819 unsigned int num_pc_event_checks = 0;
820
821 /* Handle PC events on instructions */
822 Addr oldPC;
823 do {
824 oldPC = thread->instAddr();
825 cpu.system->pcEventQueue.service(thread);
826 num_pc_event_checks++;
827 } while (oldPC != thread->instAddr());
828
829 if (num_pc_event_checks > 1) {
830 DPRINTF(PCEvent, "Acting on PC Event to PC: %s\n",
831 thread->pcState());
832 }
833
834 return num_pc_event_checks > 1;
835}
836
837void
838Execute::doInstCommitAccounting(MinorDynInstPtr inst)
839{
840 assert(!inst->isFault());
841
842 MinorThread *thread = cpu.threads[inst->id.threadId];
843
844 /* Increment the many and various inst and op counts in the
845 * thread and system */
846 if (!inst->staticInst->isMicroop() || inst->staticInst->isLastMicroop())
847 {
848 thread->numInst++;
849 thread->numInsts++;
850 cpu.stats.numInsts++;
851 cpu.system->totalNumInsts++;
852
853 /* Act on events related to instruction counts */
854 cpu.comInstEventQueue[inst->id.threadId]->serviceEvents(thread->numInst);
855 cpu.system->instEventQueue.serviceEvents(cpu.system->totalNumInsts);
856 }
857 thread->numOp++;
858 thread->numOps++;
859 cpu.stats.numOps++;
860 cpu.stats.committedInstType[inst->id.threadId]
861 [inst->staticInst->opClass()]++;
862
863 /* Set the CP SeqNum to the numOps commit number */
864 if (inst->traceData)
865 inst->traceData->setCPSeq(thread->numOp);
866
867 cpu.probeInstCommit(inst->staticInst);
868}
869
870bool
871Execute::commitInst(MinorDynInstPtr inst, bool early_memory_issue,
872 BranchData &branch, Fault &fault, bool &committed,
873 bool &completed_mem_issue)
874{
875 ThreadID thread_id = inst->id.threadId;
876 ThreadContext *thread = cpu.getContext(thread_id);
877
878 bool completed_inst = true;
879 fault = NoFault;
880
881 /* Is the thread for this instruction suspended? In that case, just
882 * stall as long as there are no pending interrupts */
883 if (thread->status() == ThreadContext::Suspended &&
884 !isInterrupted(thread_id))
885 {
886 panic("We should never hit the case where we try to commit from a "
887 "suspended thread as the streamSeqNum should not match");
888 } else if (inst->isFault()) {
889 ExecContext context(cpu, *cpu.threads[thread_id], *this, inst);
890
891 DPRINTF(MinorExecute, "Fault inst reached Execute: %s\n",
892 inst->fault->name());
893
894 fault = inst->fault;
895 inst->fault->invoke(thread, NULL);
896
897 tryToBranch(inst, fault, branch);
898 } else if (inst->staticInst->isMemRef()) {
899 /* Memory accesses are executed in two parts:
900 * executeMemRefInst -- calculates the EA and issues the access
901 * to memory. This is done here.
902 * handleMemResponse -- handles the response packet, done by
903 * Execute::commit
904 *
905 * While the memory access is in its FU, the EA is being
906 * calculated. At the end of the FU, when it is ready to
907 * 'commit' (in this function), the access is presented to the
908 * memory queues. When a response comes back from memory,
909 * Execute::commit will commit it.
910 */
911 bool predicate_passed = false;
912 bool completed_mem_inst = executeMemRefInst(inst, branch,
913 predicate_passed, fault);
914
915 if (completed_mem_inst && fault != NoFault) {
916 if (early_memory_issue) {
917 DPRINTF(MinorExecute, "Fault in early executing inst: %s\n",
918 fault->name());
919 /* Don't execute the fault, just stall the instruction
920 * until it gets to the head of inFlightInsts */
921 inst->canEarlyIssue = false;
922 /* Not completed as we'll come here again to pick up
923 * the fault when we get to the end of the FU */
924 completed_inst = false;
925 } else {
926 DPRINTF(MinorExecute, "Fault in execute: %s\n",
927 fault->name());
928 fault->invoke(thread, NULL);
929
930 tryToBranch(inst, fault, branch);
931 completed_inst = true;
932 }
933 } else {
934 completed_inst = completed_mem_inst;
935 }
936 completed_mem_issue = completed_inst;
937 } else if (inst->isInst() && inst->staticInst->isMemBarrier() &&
938 !lsq.canPushIntoStoreBuffer())
939 {
940 DPRINTF(MinorExecute, "Can't commit data barrier inst: %s yet as"
941 " there isn't space in the store buffer\n", *inst);
942
943 completed_inst = false;
944 } else if (inst->isInst() && inst->staticInst->isQuiesce()
945 && !branch.isBubble()){
946 /* This instruction can suspend, need to be able to communicate
947 * backwards, so no other branches may evaluate this cycle*/
948 completed_inst = false;
949 } else {
950 ExecContext context(cpu, *cpu.threads[thread_id], *this, inst);
951
952 DPRINTF(MinorExecute, "Committing inst: %s\n", *inst);
953
954 fault = inst->staticInst->execute(&context,
955 inst->traceData);
956
957 /* Set the predicate for tracing and dump */
958 if (inst->traceData)
959 inst->traceData->setPredicate(context.readPredicate());
960
961 committed = true;
962
963 if (fault != NoFault) {
964 DPRINTF(MinorExecute, "Fault in execute of inst: %s fault: %s\n",
965 *inst, fault->name());
966 fault->invoke(thread, inst->staticInst);
967 }
968
969 doInstCommitAccounting(inst);
970 tryToBranch(inst, fault, branch);
971 }
972
973 if (completed_inst) {
974 /* Keep a copy of this instruction's predictionSeqNum just in case
975 * we need to issue a branch without an instruction (such as an
976 * interrupt) */
977 executeInfo[thread_id].lastPredictionSeqNum = inst->id.predictionSeqNum;
978
979 /* Check to see if this instruction suspended the current thread. */
980 if (!inst->isFault() &&
981 thread->status() == ThreadContext::Suspended &&
982 branch.isBubble() && /* It didn't branch too */
983 !isInterrupted(thread_id)) /* Don't suspend if we have
984 interrupts */
985 {
986 TheISA::PCState resume_pc = cpu.getContext(thread_id)->pcState();
987
988 assert(resume_pc.microPC() == 0);
989
990 DPRINTF(MinorInterrupt, "Suspending thread: %d from Execute"
991 " inst: %s\n", thread_id, *inst);
992
993 cpu.stats.numFetchSuspends++;
994
995 updateBranchData(thread_id, BranchData::SuspendThread, inst,
996 resume_pc, branch);
997 }
998 }
999
1000 return completed_inst;
1001}
1002
1003void
1004Execute::commit(ThreadID thread_id, bool only_commit_microops, bool discard,
1005 BranchData &branch)
1006{
1007 Fault fault = NoFault;
1008 Cycles now = cpu.curCycle();
1009 ExecuteThreadInfo &ex_info = executeInfo[thread_id];
1010
1011 /**
1012 * Try and execute as many instructions from the end of FU pipelines as
1013 * possible. This *doesn't* include actually advancing the pipelines.
1014 *
1015 * We do this by looping on the front of the inFlightInsts queue for as
1016 * long as we can find the desired instruction at the end of the
1017 * functional unit it was issued to without seeing a branch or a fault.
1018 * In this function, these terms are used:
1019 * complete -- The instruction has finished its passage through
1020 * its functional unit and its fate has been decided
1021 * (committed, discarded, issued to the memory system)
1022 * commit -- The instruction is complete(d), not discarded and has
1023 * its effects applied to the CPU state
1024 * discard(ed) -- The instruction is complete but not committed
1025 * as its streamSeqNum disagrees with the current
1026 * Execute::streamSeqNum
1027 *
1028 * Commits are also possible from two other places:
1029 *
1030 * 1) Responses returning from the LSQ
1031 * 2) Mem ops issued to the LSQ ('committed' from the FUs) earlier
1032 * than their position in the inFlightInsts queue, but after all
1033 * their dependencies are resolved.
1034 */
1035
1036 /* Has an instruction been completed? Once this becomes false, we stop
1037 * trying to complete instructions. */
1038 bool completed_inst = true;
1039
1040 /* Number of insts committed this cycle to check against commitLimit */
1041 unsigned int num_insts_committed = 0;
1042
1043 /* Number of memory access instructions committed to check against
1044 * memCommitLimit */
1045 unsigned int num_mem_refs_committed = 0;
1046
1047 if (only_commit_microops && !ex_info.inFlightInsts->empty()) {
1048 DPRINTF(MinorInterrupt, "Only commit microops %s %d\n",
1049 *(ex_info.inFlightInsts->front().inst),
1050 ex_info.lastCommitWasEndOfMacroop);
1051 }
1052
1053 while (!ex_info.inFlightInsts->empty() && /* Some more instructions to process */
1054 !branch.isStreamChange() && /* No real branch */
1055 fault == NoFault && /* No faults */
1056 completed_inst && /* Still finding instructions to execute */
1057 num_insts_committed != commitLimit /* Not reached commit limit */
1058 )
1059 {
1060 if (only_commit_microops) {
1061 DPRINTF(MinorInterrupt, "Committing tail of insts before"
1062 " interrupt: %s\n",
1063 *(ex_info.inFlightInsts->front().inst));
1064 }
1065
1066 QueuedInst *head_inflight_inst = &(ex_info.inFlightInsts->front());
1067
1068 InstSeqNum head_exec_seq_num =
1069 head_inflight_inst->inst->id.execSeqNum;
1070
1071 /* The instruction we actually process if completed_inst
1072 * remains true to the end of the loop body.
1073 * Start by considering the the head of the in flight insts queue */
1074 MinorDynInstPtr inst = head_inflight_inst->inst;
1075
1076 bool committed_inst = false;
1077 bool discard_inst = false;
1078 bool completed_mem_ref = false;
1079 bool issued_mem_ref = false;
1080 bool early_memory_issue = false;
1081
1082 /* Must set this again to go around the loop */
1083 completed_inst = false;
1084
1085 /* If we're just completing a macroop before an interrupt or drain,
1086 * can we stil commit another microop (rather than a memory response)
1087 * without crosing into the next full instruction? */
1088 bool can_commit_insts = !ex_info.inFlightInsts->empty() &&
1089 !(only_commit_microops && ex_info.lastCommitWasEndOfMacroop);
1090
1091 /* Can we find a mem response for this inst */
1092 LSQ::LSQRequestPtr mem_response =
1093 (inst->inLSQ ? lsq.findResponse(inst) : NULL);
1094
1095 DPRINTF(MinorExecute, "Trying to commit canCommitInsts: %d\n",
1096 can_commit_insts);
1097
1098 /* Test for PC events after every instruction */
1099 if (isInbetweenInsts(thread_id) && tryPCEvents(thread_id)) {
1100 ThreadContext *thread = cpu.getContext(thread_id);
1101
1102 /* Branch as there was a change in PC */
1103 updateBranchData(thread_id, BranchData::UnpredictedBranch,
1104 MinorDynInst::bubble(), thread->pcState(), branch);
1105 } else if (mem_response &&
1106 num_mem_refs_committed < memoryCommitLimit)
1107 {
1108 /* Try to commit from the memory responses next */
1109 discard_inst = inst->id.streamSeqNum !=
1110 ex_info.streamSeqNum || discard;
1111
1112 DPRINTF(MinorExecute, "Trying to commit mem response: %s\n",
1113 *inst);
1114
1115 /* Complete or discard the response */
1116 if (discard_inst) {
1117 DPRINTF(MinorExecute, "Discarding mem inst: %s as its"
1118 " stream state was unexpected, expected: %d\n",
1119 *inst, ex_info.streamSeqNum);
1120
1121 lsq.popResponse(mem_response);
1122 } else {
1123 handleMemResponse(inst, mem_response, branch, fault);
1124 committed_inst = true;
1125 }
1126
1127 completed_mem_ref = true;
1128 completed_inst = true;
1129 } else if (can_commit_insts) {
1130 /* If true, this instruction will, subject to timing tweaks,
1131 * be considered for completion. try_to_commit flattens
1132 * the `if' tree a bit and allows other tests for inst
1133 * commit to be inserted here. */
1134 bool try_to_commit = false;
1135
1136 /* Try and issue memory ops early if they:
1137 * - Can push a request into the LSQ
1138 * - Have reached the end of their FUs
1139 * - Have had all their dependencies satisfied
1140 * - Are from the right stream
1141 *
1142 * For any other case, leave it to the normal instruction
1143 * issue below to handle them.
1144 */
1145 if (!ex_info.inFUMemInsts->empty() && lsq.canRequest()) {
1146 DPRINTF(MinorExecute, "Trying to commit from mem FUs\n");
1147
1148 const MinorDynInstPtr head_mem_ref_inst =
1149 ex_info.inFUMemInsts->front().inst;
1150 FUPipeline *fu = funcUnits[head_mem_ref_inst->fuIndex];
1151 const MinorDynInstPtr &fu_inst = fu->front().inst;
1152
1153 /* Use this, possibly out of order, inst as the one
1154 * to 'commit'/send to the LSQ */
1155 if (!fu_inst->isBubble() &&
1156 !fu_inst->inLSQ &&
1157 fu_inst->canEarlyIssue &&
1158 ex_info.streamSeqNum == fu_inst->id.streamSeqNum &&
1159 head_exec_seq_num > fu_inst->instToWaitFor)
1160 {
1161 DPRINTF(MinorExecute, "Issuing mem ref early"
1162 " inst: %s instToWaitFor: %d\n",
1163 *(fu_inst), fu_inst->instToWaitFor);
1164
1165 inst = fu_inst;
1166 try_to_commit = true;
1167 early_memory_issue = true;
1168 completed_inst = true;
1169 }
1170 }
1171
1172 /* Try and commit FU-less insts */
1173 if (!completed_inst && inst->isNoCostInst()) {
1174 DPRINTF(MinorExecute, "Committing no cost inst: %s", *inst);
1175
1176 try_to_commit = true;
1177 completed_inst = true;
1178 }
1179
1180 /* Try to issue from the ends of FUs and the inFlightInsts
1181 * queue */
1182 if (!completed_inst && !inst->inLSQ) {
1183 DPRINTF(MinorExecute, "Trying to commit from FUs\n");
1184
1185 /* Try to commit from a functional unit */
1186 /* Is the head inst of the expected inst's FU actually the
1187 * expected inst? */
1188 QueuedInst &fu_inst =
1189 funcUnits[inst->fuIndex]->front();
1190 InstSeqNum fu_inst_seq_num = fu_inst.inst->id.execSeqNum;
1191
1192 if (fu_inst.inst->isBubble()) {
1193 /* No instruction ready */
1194 completed_inst = false;
1195 } else if (fu_inst_seq_num != head_exec_seq_num) {
1196 /* Past instruction: we must have already executed it
1197 * in the same cycle and so the head inst isn't
1198 * actually at the end of its pipeline
1199 * Future instruction: handled above and only for
1200 * mem refs on their way to the LSQ */
1201 } else if (fu_inst.inst->id == inst->id) {
1202 /* All instructions can be committed if they have the
1203 * right execSeqNum and there are no in-flight
1204 * mem insts before us */
1205 try_to_commit = true;
1206 completed_inst = true;
1207 }
1208 }
1209
1210 if (try_to_commit) {
1211 discard_inst = inst->id.streamSeqNum !=
1212 ex_info.streamSeqNum || discard;
1213
1214 /* Is this instruction discardable as its streamSeqNum
1215 * doesn't match? */
1216 if (!discard_inst) {
1217 /* Try to commit or discard a non-memory instruction.
1218 * Memory ops are actually 'committed' from this FUs
1219 * and 'issued' into the memory system so we need to
1220 * account for them later (commit_was_mem_issue gets
1221 * set) */
1222 if (inst->extraCommitDelayExpr) {
1223 DPRINTF(MinorExecute, "Evaluating expression for"
1224 " extra commit delay inst: %s\n", *inst);
1225
1226 ThreadContext *thread = cpu.getContext(thread_id);
1227
1228 TimingExprEvalContext context(inst->staticInst,
1229 thread, NULL);
1230
1231 uint64_t extra_delay = inst->extraCommitDelayExpr->
1232 eval(context);
1233
1234 DPRINTF(MinorExecute, "Extra commit delay expr"
1235 " result: %d\n", extra_delay);
1236
1237 if (extra_delay < 128) {
1238 inst->extraCommitDelay += Cycles(extra_delay);
1239 } else {
1240 DPRINTF(MinorExecute, "Extra commit delay was"
1241 " very long: %d\n", extra_delay);
1242 }
1243 inst->extraCommitDelayExpr = NULL;
1244 }
1245
1246 /* Move the extraCommitDelay from the instruction
1247 * into the minimumCommitCycle */
1248 if (inst->extraCommitDelay != Cycles(0)) {
1249 inst->minimumCommitCycle = cpu.curCycle() +
1250 inst->extraCommitDelay;
1251 inst->extraCommitDelay = Cycles(0);
1252 }
1253
1254 /* @todo Think about making lastMemBarrier be
1255 * MAX_UINT_64 to avoid using 0 as a marker value */
1256 if (!inst->isFault() && inst->isMemRef() &&
1257 lsq.getLastMemBarrier(thread_id) <
1258 inst->id.execSeqNum &&
1259 lsq.getLastMemBarrier(thread_id) != 0)
1260 {
1261 DPRINTF(MinorExecute, "Not committing inst: %s yet"
1262 " as there are incomplete barriers in flight\n",
1263 *inst);
1264 completed_inst = false;
1265 } else if (inst->minimumCommitCycle > now) {
1266 DPRINTF(MinorExecute, "Not committing inst: %s yet"
1267 " as it wants to be stalled for %d more cycles\n",
1268 *inst, inst->minimumCommitCycle - now);
1269 completed_inst = false;
1270 } else {
1271 completed_inst = commitInst(inst,
1272 early_memory_issue, branch, fault,
1273 committed_inst, issued_mem_ref);
1274 }
1275 } else {
1276 /* Discard instruction */
1277 completed_inst = true;
1278 }
1279
1280 if (completed_inst) {
1281 /* Allow the pipeline to advance. If the FU head
1282 * instruction wasn't the inFlightInsts head
1283 * but had already been committed, it would have
1284 * unstalled the pipeline before here */
1285 if (inst->fuIndex != noCostFUIndex) {
1286 DPRINTF(MinorExecute, "Unstalling %d for inst %s\n", inst->fuIndex, inst->id);
1287 funcUnits[inst->fuIndex]->stalled = false;
1288 }
1289 }
1290 }
1291 } else {
1292 DPRINTF(MinorExecute, "No instructions to commit\n");
1293 completed_inst = false;
1294 }
1295
1296 /* All discardable instructions must also be 'completed' by now */
1297 assert(!(discard_inst && !completed_inst));
1298
1299 /* Instruction committed but was discarded due to streamSeqNum
1300 * mismatch */
1301 if (discard_inst) {
1302 DPRINTF(MinorExecute, "Discarding inst: %s as its stream"
1303 " state was unexpected, expected: %d\n",
1304 *inst, ex_info.streamSeqNum);
1305
1306 if (fault == NoFault)
1307 cpu.stats.numDiscardedOps++;
1308 }
1309
1310 /* Mark the mem inst as being in the LSQ */
1311 if (issued_mem_ref) {
1312 inst->fuIndex = 0;
1313 inst->inLSQ = true;
1314 }
1315
1316 /* Pop issued (to LSQ) and discarded mem refs from the inFUMemInsts
1317 * as they've *definitely* exited the FUs */
1318 if (completed_inst && inst->isMemRef()) {
1319 /* The MemRef could have been discarded from the FU or the memory
1320 * queue, so just check an FU instruction */
1321 if (!ex_info.inFUMemInsts->empty() &&
1322 ex_info.inFUMemInsts->front().inst == inst)
1323 {
1324 ex_info.inFUMemInsts->pop();
1325 }
1326 }
1327
1328 if (completed_inst && !(issued_mem_ref && fault == NoFault)) {
1329 /* Note that this includes discarded insts */
1330 DPRINTF(MinorExecute, "Completed inst: %s\n", *inst);
1331
1332 /* Got to the end of a full instruction? */
1333 ex_info.lastCommitWasEndOfMacroop = inst->isFault() ||
1334 inst->isLastOpInInst();
1335
1336 /* lastPredictionSeqNum is kept as a convenience to prevent its
1337 * value from changing too much on the minorview display */
1338 ex_info.lastPredictionSeqNum = inst->id.predictionSeqNum;
1339
1340 /* Finished with the inst, remove it from the inst queue and
1341 * clear its dependencies */
1342 ex_info.inFlightInsts->pop();
1343
1344 /* Complete barriers in the LSQ/move to store buffer */
1345 if (inst->isInst() && inst->staticInst->isMemBarrier()) {
1346 DPRINTF(MinorMem, "Completing memory barrier"
1347 " inst: %s committed: %d\n", *inst, committed_inst);
1348 lsq.completeMemBarrierInst(inst, committed_inst);
1349 }
1350
1351 scoreboard[thread_id].clearInstDests(inst, inst->isMemRef());
1352 }
1353
1354 /* Handle per-cycle instruction counting */
1355 if (committed_inst) {
1356 bool is_no_cost_inst = inst->isNoCostInst();
1357
1358 /* Don't show no cost instructions as having taken a commit
1359 * slot */
1360 if (DTRACE(MinorTrace) && !is_no_cost_inst)
1361 ex_info.instsBeingCommitted.insts[num_insts_committed] = inst;
1362
1363 if (!is_no_cost_inst)
1364 num_insts_committed++;
1365
1366 if (num_insts_committed == commitLimit)
1367 DPRINTF(MinorExecute, "Reached inst commit limit\n");
1368
1369 /* Re-set the time of the instruction if that's required for
1370 * tracing */
1371 if (inst->traceData) {
1372 if (setTraceTimeOnCommit)
1373 inst->traceData->setWhen(curTick());
1374 inst->traceData->dump();
1375 }
1376
1377 if (completed_mem_ref)
1378 num_mem_refs_committed++;
1379
1380 if (num_mem_refs_committed == memoryCommitLimit)
1381 DPRINTF(MinorExecute, "Reached mem ref commit limit\n");
1382 }
1383 }
1384}
1385
1386bool
1387Execute::isInbetweenInsts(ThreadID thread_id) const
1388{
1389 return executeInfo[thread_id].lastCommitWasEndOfMacroop &&
1390 !lsq.accessesInFlight();
1391}
1392
1393void
1394Execute::evaluate()
1395{
1396 if (!inp.outputWire->isBubble())
1397 inputBuffer[inp.outputWire->threadId].setTail(*inp.outputWire);
1398
1399 BranchData &branch = *out.inputWire;
1400
1401 unsigned int num_issued = 0;
1402
1403 /* Do all the cycle-wise activities for dcachePort here to potentially
1404 * free up input spaces in the LSQ's requests queue */
1405 lsq.step();
1406
1407 /* Check interrupts first. Will halt commit if interrupt found */
1408 bool interrupted = false;
1409 ThreadID interrupt_tid = checkInterrupts(branch, interrupted);
1410
1411 if (interrupt_tid != InvalidThreadID) {
1412 /* Signalling an interrupt this cycle, not issuing/committing from
1413 * any other threads */
1414 } else if (!branch.isBubble()) {
1415 /* It's important that this is here to carry Fetch1 wakeups to Fetch1
1416 * without overwriting them */
1417 DPRINTF(MinorInterrupt, "Execute skipping a cycle to allow old"
1418 " branch to complete\n");
1419 } else {
1420 ThreadID commit_tid = getCommittingThread();
1421
1422 if (commit_tid != InvalidThreadID) {
1423 ExecuteThreadInfo& commit_info = executeInfo[commit_tid];
1424
1425 DPRINTF(MinorExecute, "Attempting to commit [tid:%d]\n",
1426 commit_tid);
1427 /* commit can set stalled flags observable to issue and so *must* be
1428 * called first */
1429 if (commit_info.drainState != NotDraining) {
1430 if (commit_info.drainState == DrainCurrentInst) {
1431 /* Commit only micro-ops, don't kill anything else */
1432 commit(commit_tid, true, false, branch);
1433
1434 if (isInbetweenInsts(commit_tid))
1435 setDrainState(commit_tid, DrainHaltFetch);
1436
1437 /* Discard any generated branch */
1438 branch = BranchData::bubble();
1439 } else if (commit_info.drainState == DrainAllInsts) {
1440 /* Kill all instructions */
1441 while (getInput(commit_tid))
1442 popInput(commit_tid);
1443 commit(commit_tid, false, true, branch);
1444 }
1445 } else {
1446 /* Commit micro-ops only if interrupted. Otherwise, commit
1447 * anything you like */
1448 DPRINTF(MinorExecute, "Committing micro-ops for interrupt[tid:%d]\n",
1449 commit_tid);
1450 bool only_commit_microops = interrupted &&
1451 hasInterrupt(commit_tid);
1452 commit(commit_tid, only_commit_microops, false, branch);
1453 }
1454
1455 /* Halt fetch, but don't do it until we have the current instruction in
1456 * the bag */
1457 if (commit_info.drainState == DrainHaltFetch) {
1458 updateBranchData(commit_tid, BranchData::HaltFetch,
1459 MinorDynInst::bubble(), TheISA::PCState(0), branch);
1460
1461 cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
1462 setDrainState(commit_tid, DrainAllInsts);
1463 }
1464 }
1465 ThreadID issue_tid = getIssuingThread();
1466 /* This will issue merrily even when interrupted in the sure and
1467 * certain knowledge that the interrupt with change the stream */
1468 if (issue_tid != InvalidThreadID) {
1469 DPRINTF(MinorExecute, "Attempting to issue [tid:%d]\n",
1470 issue_tid);
1471 num_issued = issue(issue_tid);
1472 }
1473
1474 }
1475
1476 /* Run logic to step functional units + decide if we are active on the next
1477 * clock cycle */
1478 std::vector<MinorDynInstPtr> next_issuable_insts;
1479 bool can_issue_next = false;
1480
1481 for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
1482 /* Find the next issuable instruction for each thread and see if it can
1483 be issued */
1484 if (getInput(tid)) {
1485 unsigned int input_index = executeInfo[tid].inputIndex;
1486 MinorDynInstPtr inst = getInput(tid)->insts[input_index];
1487 if (inst->isFault()) {
1488 can_issue_next = true;
1489 } else if (!inst->isBubble()) {
1490 next_issuable_insts.push_back(inst);
1491 }
1492 }
1493 }
1494
1495 bool becoming_stalled = true;
1496
1497 /* Advance the pipelines and note whether they still need to be
1498 * advanced */
1499 for (unsigned int i = 0; i < numFuncUnits; i++) {
1500 FUPipeline *fu = funcUnits[i];
1501 fu->advance();
1502
1503 /* If we need to tick again, the pipeline will have been left or set
1504 * to be unstalled */
1505 if (fu->occupancy !=0 && !fu->stalled)
1506 becoming_stalled = false;
1507
1508 /* Could we possibly issue the next instruction from any thread?
1509 * This is quite an expensive test and is only used to determine
1510 * if the CPU should remain active, only run it if we aren't sure
1511 * we are active next cycle yet */
1512 for (auto inst : next_issuable_insts) {
1513 if (!fu->stalled && fu->provides(inst->staticInst->opClass()) &&
1514 scoreboard[inst->id.threadId].canInstIssue(inst,
1515 NULL, NULL, cpu.curCycle() + Cycles(1),
1516 cpu.getContext(inst->id.threadId))) {
1517 can_issue_next = true;
1518 break;
1519 }
1520 }
1521 }
1522
1523 bool head_inst_might_commit = false;
1524
1525 /* Could the head in flight insts be committed */
1526 for (auto const &info : executeInfo) {
1527 if (!info.inFlightInsts->empty()) {
1528 const QueuedInst &head_inst = info.inFlightInsts->front();
1529
1530 if (head_inst.inst->isNoCostInst()) {
1531 head_inst_might_commit = true;
1532 } else {
1533 FUPipeline *fu = funcUnits[head_inst.inst->fuIndex];
1534 if ((fu->stalled &&
1535 fu->front().inst->id == head_inst.inst->id) ||
1536 lsq.findResponse(head_inst.inst))
1537 {
1538 head_inst_might_commit = true;
1539 break;
1540 }
1541 }
1542 }
1543 }
1544
1545 DPRINTF(Activity, "Need to tick num issued insts: %s%s%s%s%s%s\n",
1546 (num_issued != 0 ? " (issued some insts)" : ""),
1547 (becoming_stalled ? "(becoming stalled)" : "(not becoming stalled)"),
1548 (can_issue_next ? " (can issued next inst)" : ""),
1549 (head_inst_might_commit ? "(head inst might commit)" : ""),
1550 (lsq.needsToTick() ? " (LSQ needs to tick)" : ""),
1551 (interrupted ? " (interrupted)" : ""));
1552
1553 bool need_to_tick =
1554 num_issued != 0 || /* Issued some insts this cycle */
1555 !becoming_stalled || /* Some FU pipelines can still move */
1556 can_issue_next || /* Can still issue a new inst */
1557 head_inst_might_commit || /* Could possible commit the next inst */
1558 lsq.needsToTick() || /* Must step the dcache port */
1559 interrupted; /* There are pending interrupts */
1560
1561 if (!need_to_tick) {
1562 DPRINTF(Activity, "The next cycle might be skippable as there are no"
1563 " advanceable FUs\n");
1564 }
1565
1566 /* Wake up if we need to tick again */
1567 if (need_to_tick)
1568 cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
1569
1570 /* Note activity of following buffer */
1571 if (!branch.isBubble())
1572 cpu.activityRecorder->activity();
1573
1574 /* Make sure the input (if any left) is pushed */
1575 if (!inp.outputWire->isBubble())
1576 inputBuffer[inp.outputWire->threadId].pushTail();
1577}
1578
1579ThreadID
1580Execute::checkInterrupts(BranchData& branch, bool& interrupted)
1581{
1582 ThreadID tid = interruptPriority;
1583 /* Evaluate interrupts in round-robin based upon service */
1584 do {
1585 /* Has an interrupt been signalled? This may not be acted on
1586 * straighaway so this is different from took_interrupt */
1587 bool thread_interrupted = false;
1588
1589 if (FullSystem && cpu.getInterruptController(tid)) {
1590 /* This is here because it seems that after drainResume the
1591 * interrupt controller isn't always set */
1592 thread_interrupted = executeInfo[tid].drainState == NotDraining &&
1593 isInterrupted(tid);
1594 interrupted = interrupted || thread_interrupted;
1595 } else {
1596 DPRINTF(MinorInterrupt, "No interrupt controller\n");
1597 }
1598 DPRINTF(MinorInterrupt, "[tid:%d] thread_interrupted?=%d isInbetweenInsts?=%d\n",
1599 tid, thread_interrupted, isInbetweenInsts(tid));
1600 /* Act on interrupts */
1601 if (thread_interrupted && isInbetweenInsts(tid)) {
1602 if (takeInterrupt(tid, branch)) {
1603 interruptPriority = tid;
1604 return tid;
1605 }
1606 } else {
1607 tid = (tid + 1) % cpu.numThreads;
1608 }
1609 } while (tid != interruptPriority);
1610
1611 return InvalidThreadID;
1612}
1613
1614bool
1615Execute::hasInterrupt(ThreadID thread_id)
1616{
1617 if (FullSystem && cpu.getInterruptController(thread_id)) {
1618 return executeInfo[thread_id].drainState == NotDraining &&
1619 isInterrupted(thread_id);
1620 }
1621
1622 return false;
1623}
1624
1625void
1626Execute::minorTrace() const
1627{
1628 std::ostringstream insts;
1629 std::ostringstream stalled;
1630
1631 executeInfo[0].instsBeingCommitted.reportData(insts);
1632 lsq.minorTrace();
1633 inputBuffer[0].minorTrace();
1634 scoreboard[0].minorTrace();
1635
1636 /* Report functional unit stalling in one string */
1637 unsigned int i = 0;
1638 while (i < numFuncUnits)
1639 {
1640 stalled << (funcUnits[i]->stalled ? '1' : 'E');
1641 i++;
1642 if (i != numFuncUnits)
1643 stalled << ',';
1644 }
1645
1646 MINORTRACE("insts=%s inputIndex=%d streamSeqNum=%d"
1647 " stalled=%s drainState=%d isInbetweenInsts=%d\n",
1648 insts.str(), executeInfo[0].inputIndex, executeInfo[0].streamSeqNum,
1649 stalled.str(), executeInfo[0].drainState, isInbetweenInsts(0));
1650
1651 std::for_each(funcUnits.begin(), funcUnits.end(),
1652 std::mem_fun(&FUPipeline::minorTrace));
1653
1654 executeInfo[0].inFlightInsts->minorTrace();
1655 executeInfo[0].inFUMemInsts->minorTrace();
1656}
1657
1658inline ThreadID
1659Execute::getCommittingThread()
1660{
1661 std::vector<ThreadID> priority_list;
1662
1663 switch (cpu.threadPolicy) {
1664 case Enums::SingleThreaded:
1665 return 0;
1666 case Enums::RoundRobin:
1667 priority_list = cpu.roundRobinPriority(commitPriority);
1668 break;
1669 case Enums::Random:
1670 priority_list = cpu.randomPriority();
1671 break;
1672 default:
1673 panic("Invalid thread policy");
1674 }
1675
1676 for (auto tid : priority_list) {
1677 ExecuteThreadInfo &ex_info = executeInfo[tid];
1678 bool can_commit_insts = !ex_info.inFlightInsts->empty();
1679 if (can_commit_insts) {
1680 QueuedInst *head_inflight_inst = &(ex_info.inFlightInsts->front());
1681 MinorDynInstPtr inst = head_inflight_inst->inst;
1682
1683 can_commit_insts = can_commit_insts &&
1684 (!inst->inLSQ || (lsq.findResponse(inst) != NULL));
1685
1686 if (!inst->inLSQ) {
1687 bool can_transfer_mem_inst = false;
1688 if (!ex_info.inFUMemInsts->empty() && lsq.canRequest()) {
1689 const MinorDynInstPtr head_mem_ref_inst =
1690 ex_info.inFUMemInsts->front().inst;
1691 FUPipeline *fu = funcUnits[head_mem_ref_inst->fuIndex];
1692 const MinorDynInstPtr &fu_inst = fu->front().inst;
1693 can_transfer_mem_inst =
1694 !fu_inst->isBubble() &&
1695 fu_inst->id.threadId == tid &&
1696 !fu_inst->inLSQ &&
1697 fu_inst->canEarlyIssue &&
1698 inst->id.execSeqNum > fu_inst->instToWaitFor;
1699 }
1700
1701 bool can_execute_fu_inst = inst->fuIndex == noCostFUIndex;
1702 if (can_commit_insts && !can_transfer_mem_inst &&
1703 inst->fuIndex != noCostFUIndex)
1704 {
1705 QueuedInst& fu_inst = funcUnits[inst->fuIndex]->front();
1706 can_execute_fu_inst = !fu_inst.inst->isBubble() &&
1707 fu_inst.inst->id == inst->id;
1708 }
1709
1710 can_commit_insts = can_commit_insts &&
1711 (can_transfer_mem_inst || can_execute_fu_inst);
1712 }
1713 }
1714
1715
1716 if (can_commit_insts) {
1717 commitPriority = tid;
1718 return tid;
1719 }
1720 }
1721
1722 return InvalidThreadID;
1723}
1724
1725inline ThreadID
1726Execute::getIssuingThread()
1727{
1728 std::vector<ThreadID> priority_list;
1729
1730 switch (cpu.threadPolicy) {
1731 case Enums::SingleThreaded:
1732 return 0;
1733 case Enums::RoundRobin:
1734 priority_list = cpu.roundRobinPriority(issuePriority);
1735 break;
1736 case Enums::Random:
1737 priority_list = cpu.randomPriority();
1738 break;
1739 default:
1740 panic("Invalid thread scheduling policy.");
1741 }
1742
1743 for (auto tid : priority_list) {
1744 if (getInput(tid)) {
1745 issuePriority = tid;
1746 return tid;
1747 }
1748 }
1749
1750 return InvalidThreadID;
1751}
1752
1753void
1754Execute::drainResume()
1755{
1756 DPRINTF(Drain, "MinorExecute drainResume\n");
1757
1758 for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
1759 setDrainState(tid, NotDraining);
1760 }
1761
1762 cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
1763}
1764
1765std::ostream &operator <<(std::ostream &os, Execute::DrainState state)
1766{
1767 switch (state)
1768 {
1769 case Execute::NotDraining:
1770 os << "NotDraining";
1771 break;
1772 case Execute::DrainCurrentInst:
1773 os << "DrainCurrentInst";
1774 break;
1775 case Execute::DrainHaltFetch:
1776 os << "DrainHaltFetch";
1777 break;
1778 case Execute::DrainAllInsts:
1779 os << "DrainAllInsts";
1780 break;
1781 default:
1782 os << "Drain-" << static_cast<int>(state);
1783 break;
1784 }
1785
1786 return os;
1787}
1788
1789void
1790Execute::setDrainState(ThreadID thread_id, DrainState state)
1791{
1792 DPRINTF(Drain, "setDrainState[%d]: %s\n", thread_id, state);
1793 executeInfo[thread_id].drainState = state;
1794}
1795
1796unsigned int
1797Execute::drain()
1798{
1799 DPRINTF(Drain, "MinorExecute drain\n");
1800
1801 for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
1802 if (executeInfo[tid].drainState == NotDraining) {
1803 cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
1804
1805 /* Go to DrainCurrentInst if we're between microops
1806 * or waiting on an unbufferable memory operation.
1807 * Otherwise we can go straight to DrainHaltFetch
1808 */
1809 if (isInbetweenInsts(tid))
1810 setDrainState(tid, DrainHaltFetch);
1811 else
1812 setDrainState(tid, DrainCurrentInst);
1813 }
1814 }
1815 return (isDrained() ? 0 : 1);
1816}
1817
1818bool
1819Execute::isDrained()
1820{
1821 if (!lsq.isDrained())
1822 return false;
1823
1824 for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
1825 if (!inputBuffer[tid].empty() ||
1826 !executeInfo[tid].inFlightInsts->empty()) {
1827
1828 return false;
1829 }
1830 }
1831
1832 return true;
1833}
1834
1835Execute::~Execute()
1836{
1837 for (unsigned int i = 0; i < numFuncUnits; i++)
1838 delete funcUnits[i];
1839
1840 for (ThreadID tid = 0; tid < cpu.numThreads; tid++)
1841 delete executeInfo[tid].inFlightInsts;
1842}
1843
1844bool
1845Execute::instIsRightStream(MinorDynInstPtr inst)
1846{
1847 return inst->id.streamSeqNum == executeInfo[inst->id.threadId].streamSeqNum;
1848}
1849
1850bool
1851Execute::instIsHeadInst(MinorDynInstPtr inst)
1852{
1853 bool ret = false;
1854
1855 if (!executeInfo[inst->id.threadId].inFlightInsts->empty())
1856 ret = executeInfo[inst->id.threadId].inFlightInsts->front().inst->id == inst->id;
1857
1858 return ret;
1859}
1860
1861MinorCPU::MinorCPUPort &
1862Execute::getDcachePort()
1863{
1864 return lsq.getDcachePort();
1865}
1866
1867}