2 * Copyright (c) 2010-2013, 2018 ARM Limited
3 * Copyright (c) 2013 Advanced Micro Devices, Inc.
6 * The license below extends only to copyright in the software and shall
7 * not be construed as granting a license to any other intellectual
8 * property including but not limited to intellectual property relating
9 * to a hardware implementation of the functionality of the software
10 * licensed hereunder. You may use the software subject to the license
11 * terms below provided that you ensure that this notice is replicated
12 * unmodified and in its entirety in all distributions of the software,
13 * modified or unmodified, in source code or in binary form.
15 * Copyright (c) 2004-2006 The Regents of The University of Michigan
16 * All rights reserved.
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions are
20 * met: redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer;
22 * redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution;
25 * neither the name of the copyright holders nor the names of its
26 * contributors may be used to endorse or promote products derived from
27 * this software without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44 #ifndef __CPU_O3_IEW_IMPL_IMPL_HH__
45 #define __CPU_O3_IEW_IMPL_IMPL_HH__
47 // @todo: Fix the instantaneous communication among all the stages within
48 // iew. There's a clear delay between issue and execute, yet backwards
49 // communication happens simultaneously.
53 #include "arch/utility.hh"
54 #include "config/the_isa.hh"
55 #include "cpu/checker/cpu.hh"
56 #include "cpu/o3/fu_pool.hh"
57 #include "cpu/o3/iew.hh"
58 #include "cpu/timebuf.hh"
59 #include "debug/Activity.hh"
60 #include "debug/Drain.hh"
61 #include "debug/IEW.hh"
62 #include "debug/O3PipeView.hh"
63 #include "params/DerivO3CPU.hh"
68 DefaultIEW<Impl>::DefaultIEW(O3CPU *_cpu, DerivO3CPUParams *params)
69 : issueToExecQueue(params->backComSize, params->forwardComSize),
71 instQueue(_cpu, this, params),
72 ldstQueue(_cpu, this, params),
73 fuPool(params->fuPool),
74 commitToIEWDelay(params->commitToIEWDelay),
75 renameToIEWDelay(params->renameToIEWDelay),
76 issueToExecuteDelay(params->issueToExecuteDelay),
77 dispatchWidth(params->dispatchWidth),
78 issueWidth(params->issueWidth),
81 wbWidth(params->wbWidth),
82 numThreads(params->numThreads)
84 if (dispatchWidth > Impl::MaxWidth)
85 fatal("dispatchWidth (%d) is larger than compiled limit (%d),\n"
86 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
87 dispatchWidth, static_cast<int>(Impl::MaxWidth));
88 if (issueWidth > Impl::MaxWidth)
89 fatal("issueWidth (%d) is larger than compiled limit (%d),\n"
90 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
91 issueWidth, static_cast<int>(Impl::MaxWidth));
92 if (wbWidth > Impl::MaxWidth)
93 fatal("wbWidth (%d) is larger than compiled limit (%d),\n"
94 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n",
95 wbWidth, static_cast<int>(Impl::MaxWidth));
101 // Setup wire to read instructions coming from issue.
102 fromIssue = issueToExecQueue.getWire(-issueToExecuteDelay);
104 // Instruction queue needs the queue between issue and execute.
105 instQueue.setIssueToExecuteQueue(&issueToExecQueue);
107 for (ThreadID tid = 0; tid < Impl::MaxThreads; tid++) {
108 dispatchStatus[tid] = Running;
109 fetchRedirect[tid] = false;
112 updateLSQNextCycle = false;
114 skidBufferMax = (renameToIEWDelay + 1) * params->renameWidth;
117 template <class Impl>
119 DefaultIEW<Impl>::name() const
121 return cpu->name() + ".iew";
124 template <class Impl>
126 DefaultIEW<Impl>::regProbePoints()
128 ppDispatch = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Dispatch");
129 ppMispredict = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Mispredict");
131 * Probe point with dynamic instruction as the argument used to probe when
132 * an instruction starts to execute.
134 ppExecute = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(),
137 * Probe point with dynamic instruction as the argument used to probe when
138 * an instruction execution completes and it is marked ready to commit.
140 ppToCommit = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(),
144 template <class Impl>
146 DefaultIEW<Impl>::regStats()
148 using namespace Stats;
150 instQueue.regStats();
151 ldstQueue.regStats();
154 .name(name() + ".iewIdleCycles")
155 .desc("Number of cycles IEW is idle");
158 .name(name() + ".iewSquashCycles")
159 .desc("Number of cycles IEW is squashing");
162 .name(name() + ".iewBlockCycles")
163 .desc("Number of cycles IEW is blocking");
166 .name(name() + ".iewUnblockCycles")
167 .desc("Number of cycles IEW is unblocking");
170 .name(name() + ".iewDispatchedInsts")
171 .desc("Number of instructions dispatched to IQ");
174 .name(name() + ".iewDispSquashedInsts")
175 .desc("Number of squashed instructions skipped by dispatch");
178 .name(name() + ".iewDispLoadInsts")
179 .desc("Number of dispatched load instructions");
182 .name(name() + ".iewDispStoreInsts")
183 .desc("Number of dispatched store instructions");
186 .name(name() + ".iewDispNonSpecInsts")
187 .desc("Number of dispatched non-speculative instructions");
190 .name(name() + ".iewIQFullEvents")
191 .desc("Number of times the IQ has become full, causing a stall");
194 .name(name() + ".iewLSQFullEvents")
195 .desc("Number of times the LSQ has become full, causing a stall");
197 memOrderViolationEvents
198 .name(name() + ".memOrderViolationEvents")
199 .desc("Number of memory order violations");
201 predictedTakenIncorrect
202 .name(name() + ".predictedTakenIncorrect")
203 .desc("Number of branches that were predicted taken incorrectly");
205 predictedNotTakenIncorrect
206 .name(name() + ".predictedNotTakenIncorrect")
207 .desc("Number of branches that were predicted not taken incorrectly");
210 .name(name() + ".branchMispredicts")
211 .desc("Number of branch mispredicts detected at execute");
213 branchMispredicts = predictedTakenIncorrect + predictedNotTakenIncorrect;
216 .name(name() + ".iewExecutedInsts")
217 .desc("Number of executed instructions");
220 .init(cpu->numThreads)
221 .name(name() + ".iewExecLoadInsts")
222 .desc("Number of load instructions executed")
226 .name(name() + ".iewExecSquashedInsts")
227 .desc("Number of squashed instructions skipped in execute");
230 .init(cpu->numThreads)
231 .name(name() + ".exec_swp")
232 .desc("number of swp insts executed")
236 .init(cpu->numThreads)
237 .name(name() + ".exec_nop")
238 .desc("number of nop insts executed")
242 .init(cpu->numThreads)
243 .name(name() + ".exec_refs")
244 .desc("number of memory reference insts executed")
248 .init(cpu->numThreads)
249 .name(name() + ".exec_branches")
250 .desc("Number of branches executed")
254 .name(name() + ".exec_stores")
255 .desc("Number of stores executed")
257 iewExecStoreInsts = iewExecutedRefs - iewExecLoadInsts;
260 .name(name() + ".exec_rate")
261 .desc("Inst execution rate")
264 iewExecRate = iewExecutedInsts / cpu->numCycles;
267 .init(cpu->numThreads)
268 .name(name() + ".wb_sent")
269 .desc("cumulative count of insts sent to commit")
273 .init(cpu->numThreads)
274 .name(name() + ".wb_count")
275 .desc("cumulative count of insts written-back")
279 .init(cpu->numThreads)
280 .name(name() + ".wb_producers")
281 .desc("num instructions producing a value")
285 .init(cpu->numThreads)
286 .name(name() + ".wb_consumers")
287 .desc("num instructions consuming a value")
291 .name(name() + ".wb_fanout")
292 .desc("average fanout of values written-back")
295 wbFanout = producerInst / consumerInst;
298 .name(name() + ".wb_rate")
299 .desc("insts written-back per cycle")
301 wbRate = writebackCount / cpu->numCycles;
306 DefaultIEW<Impl>::startupStage()
308 for (ThreadID tid = 0; tid < numThreads; tid++) {
309 toRename->iewInfo[tid].usedIQ = true;
310 toRename->iewInfo[tid].freeIQEntries =
311 instQueue.numFreeEntries(tid);
313 toRename->iewInfo[tid].usedLSQ = true;
314 toRename->iewInfo[tid].freeLQEntries = ldstQueue.numFreeLoadEntries(tid);
315 toRename->iewInfo[tid].freeSQEntries = ldstQueue.numFreeStoreEntries(tid);
318 // Initialize the checker's dcache port here
320 cpu->checker->setDcachePort(&cpu->getDataPort());
323 cpu->activateStage(O3CPU::IEWIdx);
328 DefaultIEW<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
332 // Setup wire to read information from time buffer, from commit.
333 fromCommit = timeBuffer->getWire(-commitToIEWDelay);
335 // Setup wire to write information back to previous stages.
336 toRename = timeBuffer->getWire(0);
338 toFetch = timeBuffer->getWire(0);
340 // Instruction queue also needs main time buffer.
341 instQueue.setTimeBuffer(tb_ptr);
346 DefaultIEW<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
348 renameQueue = rq_ptr;
350 // Setup wire to read information from rename queue.
351 fromRename = renameQueue->getWire(-renameToIEWDelay);
356 DefaultIEW<Impl>::setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr)
360 // Setup wire to write instructions to commit.
361 toCommit = iewQueue->getWire(0);
366 DefaultIEW<Impl>::setActiveThreads(list<ThreadID> *at_ptr)
368 activeThreads = at_ptr;
370 ldstQueue.setActiveThreads(at_ptr);
371 instQueue.setActiveThreads(at_ptr);
376 DefaultIEW<Impl>::setScoreboard(Scoreboard *sb_ptr)
381 template <class Impl>
383 DefaultIEW<Impl>::isDrained() const
385 bool drained = ldstQueue.isDrained() && instQueue.isDrained();
387 for (ThreadID tid = 0; tid < numThreads; tid++) {
388 if (!insts[tid].empty()) {
389 DPRINTF(Drain, "%i: Insts not empty.\n", tid);
392 if (!skidBuffer[tid].empty()) {
393 DPRINTF(Drain, "%i: Skid buffer not empty.\n", tid);
396 drained = drained && dispatchStatus[tid] == Running;
399 // Also check the FU pool as instructions are "stored" in FU
400 // completion events until they are done and not accounted for
402 if (drained && !fuPool->isDrained()) {
403 DPRINTF(Drain, "FU pool still busy.\n");
410 template <class Impl>
412 DefaultIEW<Impl>::drainSanityCheck() const
416 instQueue.drainSanityCheck();
417 ldstQueue.drainSanityCheck();
420 template <class Impl>
422 DefaultIEW<Impl>::takeOverFrom()
429 instQueue.takeOverFrom();
430 ldstQueue.takeOverFrom();
431 fuPool->takeOverFrom();
434 cpu->activityThisCycle();
436 for (ThreadID tid = 0; tid < numThreads; tid++) {
437 dispatchStatus[tid] = Running;
438 fetchRedirect[tid] = false;
441 updateLSQNextCycle = false;
443 for (int i = 0; i < issueToExecQueue.getSize(); ++i) {
444 issueToExecQueue.advance();
450 DefaultIEW<Impl>::squash(ThreadID tid)
452 DPRINTF(IEW, "[tid:%i]: Squashing all instructions.\n", tid);
454 // Tell the IQ to start squashing.
455 instQueue.squash(tid);
457 // Tell the LDSTQ to start squashing.
458 ldstQueue.squash(fromCommit->commitInfo[tid].doneSeqNum, tid);
459 updatedQueues = true;
461 // Clear the skid buffer in case it has any data in it.
462 DPRINTF(IEW, "[tid:%i]: Removing skidbuffer instructions until [sn:%i].\n",
463 tid, fromCommit->commitInfo[tid].doneSeqNum);
465 while (!skidBuffer[tid].empty()) {
466 if (skidBuffer[tid].front()->isLoad()) {
467 toRename->iewInfo[tid].dispatchedToLQ++;
469 if (skidBuffer[tid].front()->isStore()) {
470 toRename->iewInfo[tid].dispatchedToSQ++;
473 toRename->iewInfo[tid].dispatched++;
475 skidBuffer[tid].pop();
478 emptyRenameInsts(tid);
483 DefaultIEW<Impl>::squashDueToBranch(const DynInstPtr& inst, ThreadID tid)
485 DPRINTF(IEW, "[tid:%i]: Squashing from a specific instruction, PC: %s "
486 "[sn:%i].\n", tid, inst->pcState(), inst->seqNum);
488 if (!toCommit->squash[tid] ||
489 inst->seqNum < toCommit->squashedSeqNum[tid]) {
490 toCommit->squash[tid] = true;
491 toCommit->squashedSeqNum[tid] = inst->seqNum;
492 toCommit->branchTaken[tid] = inst->pcState().branching();
494 TheISA::PCState pc = inst->pcState();
495 TheISA::advancePC(pc, inst->staticInst);
497 toCommit->pc[tid] = pc;
498 toCommit->mispredictInst[tid] = inst;
499 toCommit->includeSquashInst[tid] = false;
501 wroteToTimeBuffer = true;
508 DefaultIEW<Impl>::squashDueToMemOrder(const DynInstPtr& inst, ThreadID tid)
510 DPRINTF(IEW, "[tid:%i]: Memory violation, squashing violator and younger "
511 "insts, PC: %s [sn:%i].\n", tid, inst->pcState(), inst->seqNum);
512 // Need to include inst->seqNum in the following comparison to cover the
513 // corner case when a branch misprediction and a memory violation for the
514 // same instruction (e.g. load PC) are detected in the same cycle. In this
515 // case the memory violator should take precedence over the branch
516 // misprediction because it requires the violator itself to be included in
518 if (!toCommit->squash[tid] ||
519 inst->seqNum <= toCommit->squashedSeqNum[tid]) {
520 toCommit->squash[tid] = true;
522 toCommit->squashedSeqNum[tid] = inst->seqNum;
523 toCommit->pc[tid] = inst->pcState();
524 toCommit->mispredictInst[tid] = NULL;
526 // Must include the memory violator in the squash.
527 toCommit->includeSquashInst[tid] = true;
529 wroteToTimeBuffer = true;
535 DefaultIEW<Impl>::block(ThreadID tid)
537 DPRINTF(IEW, "[tid:%u]: Blocking.\n", tid);
539 if (dispatchStatus[tid] != Blocked &&
540 dispatchStatus[tid] != Unblocking) {
541 toRename->iewBlock[tid] = true;
542 wroteToTimeBuffer = true;
545 // Add the current inputs to the skid buffer so they can be
546 // reprocessed when this stage unblocks.
549 dispatchStatus[tid] = Blocked;
554 DefaultIEW<Impl>::unblock(ThreadID tid)
556 DPRINTF(IEW, "[tid:%i]: Reading instructions out of the skid "
557 "buffer %u.\n",tid, tid);
559 // If the skid bufffer is empty, signal back to previous stages to unblock.
560 // Also switch status to running.
561 if (skidBuffer[tid].empty()) {
562 toRename->iewUnblock[tid] = true;
563 wroteToTimeBuffer = true;
564 DPRINTF(IEW, "[tid:%i]: Done unblocking.\n",tid);
565 dispatchStatus[tid] = Running;
571 DefaultIEW<Impl>::wakeDependents(const DynInstPtr& inst)
573 instQueue.wakeDependents(inst);
578 DefaultIEW<Impl>::rescheduleMemInst(const DynInstPtr& inst)
580 instQueue.rescheduleMemInst(inst);
585 DefaultIEW<Impl>::replayMemInst(const DynInstPtr& inst)
587 instQueue.replayMemInst(inst);
592 DefaultIEW<Impl>::blockMemInst(const DynInstPtr& inst)
594 instQueue.blockMemInst(inst);
599 DefaultIEW<Impl>::cacheUnblocked()
601 instQueue.cacheUnblocked();
606 DefaultIEW<Impl>::instToCommit(const DynInstPtr& inst)
608 // This function should not be called after writebackInsts in a
609 // single cycle. That will cause problems with an instruction
610 // being added to the queue to commit without being processed by
611 // writebackInsts prior to being sent to commit.
613 // First check the time slot that this instruction will write
614 // to. If there are free write ports at the time, then go ahead
615 // and write the instruction to that time. If there are not,
616 // keep looking back to see where's the first time there's a
618 while ((*iewQueue)[wbCycle].insts[wbNumInst]) {
620 if (wbNumInst == wbWidth) {
626 DPRINTF(IEW, "Current wb cycle: %i, width: %i, numInst: %i\nwbActual:%i\n",
627 wbCycle, wbWidth, wbNumInst, wbCycle * wbWidth + wbNumInst);
628 // Add finished instruction to queue to commit.
629 (*iewQueue)[wbCycle].insts[wbNumInst] = inst;
630 (*iewQueue)[wbCycle].size++;
633 template <class Impl>
635 DefaultIEW<Impl>::validInstsFromRename()
637 unsigned inst_count = 0;
639 for (int i=0; i<fromRename->size; i++) {
640 if (!fromRename->insts[i]->isSquashed())
649 DefaultIEW<Impl>::skidInsert(ThreadID tid)
651 DynInstPtr inst = NULL;
653 while (!insts[tid].empty()) {
654 inst = insts[tid].front();
658 DPRINTF(IEW,"[tid:%i]: Inserting [sn:%lli] PC:%s into "
659 "dispatch skidBuffer %i\n",tid, inst->seqNum,
660 inst->pcState(),tid);
662 skidBuffer[tid].push(inst);
665 assert(skidBuffer[tid].size() <= skidBufferMax &&
666 "Skidbuffer Exceeded Max Size");
671 DefaultIEW<Impl>::skidCount()
675 list<ThreadID>::iterator threads = activeThreads->begin();
676 list<ThreadID>::iterator end = activeThreads->end();
678 while (threads != end) {
679 ThreadID tid = *threads++;
680 unsigned thread_count = skidBuffer[tid].size();
681 if (max < thread_count)
690 DefaultIEW<Impl>::skidsEmpty()
692 list<ThreadID>::iterator threads = activeThreads->begin();
693 list<ThreadID>::iterator end = activeThreads->end();
695 while (threads != end) {
696 ThreadID tid = *threads++;
698 if (!skidBuffer[tid].empty())
705 template <class Impl>
707 DefaultIEW<Impl>::updateStatus()
709 bool any_unblocking = false;
711 list<ThreadID>::iterator threads = activeThreads->begin();
712 list<ThreadID>::iterator end = activeThreads->end();
714 while (threads != end) {
715 ThreadID tid = *threads++;
717 if (dispatchStatus[tid] == Unblocking) {
718 any_unblocking = true;
723 // If there are no ready instructions waiting to be scheduled by the IQ,
724 // and there's no stores waiting to write back, and dispatch is not
725 // unblocking, then there is no internal activity for the IEW stage.
726 instQueue.intInstQueueReads++;
727 if (_status == Active && !instQueue.hasReadyInsts() &&
728 !ldstQueue.willWB() && !any_unblocking) {
729 DPRINTF(IEW, "IEW switching to idle\n");
734 } else if (_status == Inactive && (instQueue.hasReadyInsts() ||
735 ldstQueue.willWB() ||
737 // Otherwise there is internal activity. Set to active.
738 DPRINTF(IEW, "IEW switching to active\n");
746 template <class Impl>
748 DefaultIEW<Impl>::checkStall(ThreadID tid)
752 if (fromCommit->commitInfo[tid].robSquashing) {
753 DPRINTF(IEW,"[tid:%i]: Stall from Commit stage detected.\n",tid);
755 } else if (instQueue.isFull(tid)) {
756 DPRINTF(IEW,"[tid:%i]: Stall: IQ is full.\n",tid);
763 template <class Impl>
765 DefaultIEW<Impl>::checkSignalsAndUpdate(ThreadID tid)
767 // Check if there's a squash signal, squash if there is
768 // Check stall signals, block if there is.
769 // If status was Blocked
770 // if so then go to unblocking
771 // If status was Squashing
772 // check if squashing is not high. Switch to running this cycle.
774 if (fromCommit->commitInfo[tid].squash) {
777 if (dispatchStatus[tid] == Blocked ||
778 dispatchStatus[tid] == Unblocking) {
779 toRename->iewUnblock[tid] = true;
780 wroteToTimeBuffer = true;
783 dispatchStatus[tid] = Squashing;
784 fetchRedirect[tid] = false;
788 if (fromCommit->commitInfo[tid].robSquashing) {
789 DPRINTF(IEW, "[tid:%i]: ROB is still squashing.\n", tid);
791 dispatchStatus[tid] = Squashing;
792 emptyRenameInsts(tid);
793 wroteToTimeBuffer = true;
796 if (checkStall(tid)) {
798 dispatchStatus[tid] = Blocked;
802 if (dispatchStatus[tid] == Blocked) {
803 // Status from previous cycle was blocked, but there are no more stall
804 // conditions. Switch over to unblocking.
805 DPRINTF(IEW, "[tid:%i]: Done blocking, switching to unblocking.\n",
808 dispatchStatus[tid] = Unblocking;
815 if (dispatchStatus[tid] == Squashing) {
816 // Switch status to running if rename isn't being told to block or
817 // squash this cycle.
818 DPRINTF(IEW, "[tid:%i]: Done squashing, switching to running.\n",
821 dispatchStatus[tid] = Running;
827 template <class Impl>
829 DefaultIEW<Impl>::sortInsts()
831 int insts_from_rename = fromRename->size;
833 for (ThreadID tid = 0; tid < numThreads; tid++)
834 assert(insts[tid].empty());
836 for (int i = 0; i < insts_from_rename; ++i) {
837 insts[fromRename->insts[i]->threadNumber].push(fromRename->insts[i]);
841 template <class Impl>
843 DefaultIEW<Impl>::emptyRenameInsts(ThreadID tid)
845 DPRINTF(IEW, "[tid:%i]: Removing incoming rename instructions\n", tid);
847 while (!insts[tid].empty()) {
849 if (insts[tid].front()->isLoad()) {
850 toRename->iewInfo[tid].dispatchedToLQ++;
852 if (insts[tid].front()->isStore()) {
853 toRename->iewInfo[tid].dispatchedToSQ++;
856 toRename->iewInfo[tid].dispatched++;
862 template <class Impl>
864 DefaultIEW<Impl>::wakeCPU()
869 template <class Impl>
871 DefaultIEW<Impl>::activityThisCycle()
873 DPRINTF(Activity, "Activity this cycle.\n");
874 cpu->activityThisCycle();
877 template <class Impl>
879 DefaultIEW<Impl>::activateStage()
881 DPRINTF(Activity, "Activating stage.\n");
882 cpu->activateStage(O3CPU::IEWIdx);
885 template <class Impl>
887 DefaultIEW<Impl>::deactivateStage()
889 DPRINTF(Activity, "Deactivating stage.\n");
890 cpu->deactivateStage(O3CPU::IEWIdx);
895 DefaultIEW<Impl>::dispatch(ThreadID tid)
897 // If status is Running or idle,
898 // call dispatchInsts()
899 // If status is Unblocking,
900 // buffer any instructions coming from rename
901 // continue trying to empty skid buffer
902 // check if stall conditions have passed
904 if (dispatchStatus[tid] == Blocked) {
907 } else if (dispatchStatus[tid] == Squashing) {
911 // Dispatch should try to dispatch as many instructions as its bandwidth
912 // will allow, as long as it is not currently blocked.
913 if (dispatchStatus[tid] == Running ||
914 dispatchStatus[tid] == Idle) {
915 DPRINTF(IEW, "[tid:%i] Not blocked, so attempting to run "
919 } else if (dispatchStatus[tid] == Unblocking) {
920 // Make sure that the skid buffer has something in it if the
921 // status is unblocking.
922 assert(!skidsEmpty());
924 // If the status was unblocking, then instructions from the skid
925 // buffer were used. Remove those instructions and handle
926 // the rest of unblocking.
931 if (validInstsFromRename()) {
932 // Add the current inputs to the skid buffer so they can be
933 // reprocessed when this stage unblocks.
941 template <class Impl>
943 DefaultIEW<Impl>::dispatchInsts(ThreadID tid)
945 // Obtain instructions from skid buffer if unblocking, or queue from rename
947 std::queue<DynInstPtr> &insts_to_dispatch =
948 dispatchStatus[tid] == Unblocking ?
949 skidBuffer[tid] : insts[tid];
951 int insts_to_add = insts_to_dispatch.size();
954 bool add_to_iq = false;
955 int dis_num_inst = 0;
957 // Loop through the instructions, putting them in the instruction
959 for ( ; dis_num_inst < insts_to_add &&
960 dis_num_inst < dispatchWidth;
963 inst = insts_to_dispatch.front();
965 if (dispatchStatus[tid] == Unblocking) {
966 DPRINTF(IEW, "[tid:%i]: Issue: Examining instruction from skid "
970 // Make sure there's a valid instruction there.
973 DPRINTF(IEW, "[tid:%i]: Issue: Adding PC %s [sn:%lli] [tid:%i] to "
975 tid, inst->pcState(), inst->seqNum, inst->threadNumber);
977 // Be sure to mark these instructions as ready so that the
978 // commit stage can go ahead and execute them, and mark
979 // them as issued so the IQ doesn't reprocess them.
981 // Check for squashed instructions.
982 if (inst->isSquashed()) {
983 DPRINTF(IEW, "[tid:%i]: Issue: Squashed instruction encountered, "
984 "not adding to IQ.\n", tid);
986 ++iewDispSquashedInsts;
988 insts_to_dispatch.pop();
990 //Tell Rename That An Instruction has been processed
991 if (inst->isLoad()) {
992 toRename->iewInfo[tid].dispatchedToLQ++;
994 if (inst->isStore()) {
995 toRename->iewInfo[tid].dispatchedToSQ++;
998 toRename->iewInfo[tid].dispatched++;
1003 // Check for full conditions.
1004 if (instQueue.isFull(tid)) {
1005 DPRINTF(IEW, "[tid:%i]: Issue: IQ has become full.\n", tid);
1007 // Call function to start blocking.
1010 // Set unblock to false. Special case where we are using
1011 // skidbuffer (unblocking) instructions but then we still
1012 // get full in the IQ.
1013 toRename->iewUnblock[tid] = false;
1019 // Check LSQ if inst is LD/ST
1020 if ((inst->isLoad() && ldstQueue.lqFull(tid)) ||
1021 (inst->isStore() && ldstQueue.sqFull(tid))) {
1022 DPRINTF(IEW, "[tid:%i]: Issue: %s has become full.\n",tid,
1023 inst->isLoad() ? "LQ" : "SQ");
1025 // Call function to start blocking.
1028 // Set unblock to false. Special case where we are using
1029 // skidbuffer (unblocking) instructions but then we still
1030 // get full in the IQ.
1031 toRename->iewUnblock[tid] = false;
1037 // Otherwise issue the instruction just fine.
1038 if (inst->isLoad()) {
1039 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1040 "encountered, adding to LSQ.\n", tid);
1042 // Reserve a spot in the load store queue for this
1044 ldstQueue.insertLoad(inst);
1050 toRename->iewInfo[tid].dispatchedToLQ++;
1051 } else if (inst->isStore()) {
1052 DPRINTF(IEW, "[tid:%i]: Issue: Memory instruction "
1053 "encountered, adding to LSQ.\n", tid);
1055 ldstQueue.insertStore(inst);
1057 ++iewDispStoreInsts;
1059 if (inst->isStoreConditional()) {
1060 // Store conditionals need to be set as "canCommit()"
1061 // so that commit can process them when they reach the
1063 // @todo: This is somewhat specific to Alpha.
1064 inst->setCanCommit();
1065 instQueue.insertNonSpec(inst);
1068 ++iewDispNonSpecInsts;
1073 toRename->iewInfo[tid].dispatchedToSQ++;
1074 } else if (inst->isMemBarrier() || inst->isWriteBarrier()) {
1075 // Same as non-speculative stores.
1076 inst->setCanCommit();
1077 instQueue.insertBarrier(inst);
1079 } else if (inst->isNop()) {
1080 DPRINTF(IEW, "[tid:%i]: Issue: Nop instruction encountered, "
1081 "skipping.\n", tid);
1084 inst->setExecuted();
1085 inst->setCanCommit();
1087 instQueue.recordProducer(inst);
1089 iewExecutedNop[tid]++;
1093 assert(!inst->isExecuted());
1097 if (add_to_iq && inst->isNonSpeculative()) {
1098 DPRINTF(IEW, "[tid:%i]: Issue: Nonspeculative instruction "
1099 "encountered, skipping.\n", tid);
1101 // Same as non-speculative stores.
1102 inst->setCanCommit();
1104 // Specifically insert it as nonspeculative.
1105 instQueue.insertNonSpec(inst);
1107 ++iewDispNonSpecInsts;
1112 // If the instruction queue is not full, then add the
1115 instQueue.insert(inst);
1118 insts_to_dispatch.pop();
1120 toRename->iewInfo[tid].dispatched++;
1122 ++iewDispatchedInsts;
1125 inst->dispatchTick = curTick() - inst->fetchTick;
1127 ppDispatch->notify(inst);
1130 if (!insts_to_dispatch.empty()) {
1131 DPRINTF(IEW,"[tid:%i]: Issue: Bandwidth Full. Blocking.\n", tid);
1133 toRename->iewUnblock[tid] = false;
1136 if (dispatchStatus[tid] == Idle && dis_num_inst) {
1137 dispatchStatus[tid] = Running;
1139 updatedQueues = true;
1145 template <class Impl>
1147 DefaultIEW<Impl>::printAvailableInsts()
1151 std::cout << "Available Instructions: ";
1153 while (fromIssue->insts[inst]) {
1155 if (inst%3==0) std::cout << "\n\t";
1157 std::cout << "PC: " << fromIssue->insts[inst]->pcState()
1158 << " TN: " << fromIssue->insts[inst]->threadNumber
1159 << " SN: " << fromIssue->insts[inst]->seqNum << " | ";
1168 template <class Impl>
1170 DefaultIEW<Impl>::executeInsts()
1175 list<ThreadID>::iterator threads = activeThreads->begin();
1176 list<ThreadID>::iterator end = activeThreads->end();
1178 while (threads != end) {
1179 ThreadID tid = *threads++;
1180 fetchRedirect[tid] = false;
1183 // Uncomment this if you want to see all available instructions.
1184 // @todo This doesn't actually work anymore, we should fix it.
1185 // printAvailableInsts();
1187 // Execute/writeback any instructions that are available.
1188 int insts_to_execute = fromIssue->size;
1190 for (; inst_num < insts_to_execute;
1193 DPRINTF(IEW, "Execute: Executing instructions from IQ.\n");
1195 DynInstPtr inst = instQueue.getInstToExecute();
1197 DPRINTF(IEW, "Execute: Processing PC %s, [tid:%i] [sn:%i].\n",
1198 inst->pcState(), inst->threadNumber,inst->seqNum);
1200 // Notify potential listeners that this instruction has started
1202 ppExecute->notify(inst);
1204 // Check if the instruction is squashed; if so then skip it
1205 if (inst->isSquashed()) {
1206 DPRINTF(IEW, "Execute: Instruction was squashed. PC: %s, [tid:%i]"
1207 " [sn:%i]\n", inst->pcState(), inst->threadNumber,
1210 // Consider this instruction executed so that commit can go
1211 // ahead and retire the instruction.
1212 inst->setExecuted();
1214 // Not sure if I should set this here or just let commit try to
1215 // commit any squashed instructions. I like the latter a bit more.
1216 inst->setCanCommit();
1218 ++iewExecSquashedInsts;
1223 Fault fault = NoFault;
1225 // Execute instruction.
1226 // Note that if the instruction faults, it will be handled
1227 // at the commit stage.
1228 if (inst->isMemRef()) {
1229 DPRINTF(IEW, "Execute: Calculating address for memory "
1232 // Tell the LDSTQ to execute this instruction (if it is a load).
1233 if (inst->isLoad()) {
1234 // Loads will mark themselves as executed, and their writeback
1235 // event adds the instruction to the queue to commit
1236 fault = ldstQueue.executeLoad(inst);
1238 if (inst->isTranslationDelayed() &&
1240 // A hw page table walk is currently going on; the
1241 // instruction must be deferred.
1242 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1244 instQueue.deferMemInst(inst);
1248 if (inst->isDataPrefetch() || inst->isInstPrefetch()) {
1249 inst->fault = NoFault;
1251 } else if (inst->isStore()) {
1252 fault = ldstQueue.executeStore(inst);
1254 if (inst->isTranslationDelayed() &&
1256 // A hw page table walk is currently going on; the
1257 // instruction must be deferred.
1258 DPRINTF(IEW, "Execute: Delayed translation, deferring "
1260 instQueue.deferMemInst(inst);
1264 // If the store had a fault then it may not have a mem req
1265 if (fault != NoFault || !inst->readPredicate() ||
1266 !inst->isStoreConditional()) {
1267 // If the instruction faulted, then we need to send it along
1268 // to commit without the instruction completing.
1269 // Send this instruction to commit, also make sure iew stage
1270 // realizes there is activity.
1271 inst->setExecuted();
1273 activityThisCycle();
1276 // Store conditionals will mark themselves as
1277 // executed, and their writeback event will add the
1278 // instruction to the queue to commit.
1280 panic("Unexpected memory type!\n");
1284 // If the instruction has already faulted, then skip executing it.
1285 // Such case can happen when it faulted during ITLB translation.
1286 // If we execute the instruction (even if it's a nop) the fault
1287 // will be replaced and we will lose it.
1288 if (inst->getFault() == NoFault) {
1290 if (!inst->readPredicate())
1291 inst->forwardOldRegs();
1294 inst->setExecuted();
1299 updateExeInstStats(inst);
1301 // Check if branch prediction was correct, if not then we need
1302 // to tell commit to squash in flight instructions. Only
1303 // handle this if there hasn't already been something that
1304 // redirects fetch in this group of instructions.
1306 // This probably needs to prioritize the redirects if a different
1307 // scheduler is used. Currently the scheduler schedules the oldest
1308 // instruction first, so the branch resolution order will be correct.
1309 ThreadID tid = inst->threadNumber;
1311 if (!fetchRedirect[tid] ||
1312 !toCommit->squash[tid] ||
1313 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1315 // Prevent testing for misprediction on load instructions,
1316 // that have not been executed.
1317 bool loadNotExecuted = !inst->isExecuted() && inst->isLoad();
1319 if (inst->mispredicted() && !loadNotExecuted) {
1320 fetchRedirect[tid] = true;
1322 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1323 DPRINTF(IEW, "Predicted target was PC: %s.\n",
1324 inst->readPredTarg());
1325 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %s.\n",
1327 // If incorrect, then signal the ROB that it must be squashed.
1328 squashDueToBranch(inst, tid);
1330 ppMispredict->notify(inst);
1332 if (inst->readPredTaken()) {
1333 predictedTakenIncorrect++;
1335 predictedNotTakenIncorrect++;
1337 } else if (ldstQueue.violation(tid)) {
1338 assert(inst->isMemRef());
1339 // If there was an ordering violation, then get the
1340 // DynInst that caused the violation. Note that this
1341 // clears the violation signal.
1342 DynInstPtr violator;
1343 violator = ldstQueue.getMemDepViolator(tid);
1345 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: %s "
1346 "[sn:%lli], inst PC: %s [sn:%lli]. Addr is: %#x.\n",
1347 violator->pcState(), violator->seqNum,
1348 inst->pcState(), inst->seqNum, inst->physEffAddr);
1350 fetchRedirect[tid] = true;
1352 // Tell the instruction queue that a violation has occured.
1353 instQueue.violation(inst, violator);
1356 squashDueToMemOrder(violator, tid);
1358 ++memOrderViolationEvents;
1361 // Reset any state associated with redirects that will not
1363 if (ldstQueue.violation(tid)) {
1364 assert(inst->isMemRef());
1366 DynInstPtr violator = ldstQueue.getMemDepViolator(tid);
1368 DPRINTF(IEW, "LDSTQ detected a violation. Violator PC: "
1369 "%s, inst PC: %s. Addr is: %#x.\n",
1370 violator->pcState(), inst->pcState(),
1372 DPRINTF(IEW, "Violation will not be handled because "
1373 "already squashing\n");
1375 ++memOrderViolationEvents;
1380 // Update and record activity if we processed any instructions.
1382 if (exeStatus == Idle) {
1383 exeStatus = Running;
1386 updatedQueues = true;
1388 cpu->activityThisCycle();
1391 // Need to reset this in case a writeback event needs to write into the
1392 // iew queue. That way the writeback event will write into the correct
1393 // spot in the queue.
1398 template <class Impl>
1400 DefaultIEW<Impl>::writebackInsts()
1402 // Loop through the head of the time buffer and wake any
1403 // dependents. These instructions are about to write back. Also
1404 // mark scoreboard that this instruction is finally complete.
1405 // Either have IEW have direct access to scoreboard, or have this
1406 // as part of backwards communication.
1407 for (int inst_num = 0; inst_num < wbWidth &&
1408 toCommit->insts[inst_num]; inst_num++) {
1409 DynInstPtr inst = toCommit->insts[inst_num];
1410 ThreadID tid = inst->threadNumber;
1412 DPRINTF(IEW, "Sending instructions to commit, [sn:%lli] PC %s.\n",
1413 inst->seqNum, inst->pcState());
1415 iewInstsToCommit[tid]++;
1416 // Notify potential listeners that execution is complete for this
1418 ppToCommit->notify(inst);
1420 // Some instructions will be sent to commit without having
1421 // executed because they need commit to handle them.
1422 // E.g. Strictly ordered loads have not actually executed when they
1423 // are first sent to commit. Instead commit must tell the LSQ
1424 // when it's ready to execute the strictly ordered load.
1425 if (!inst->isSquashed() && inst->isExecuted() && inst->getFault() == NoFault) {
1426 int dependents = instQueue.wakeDependents(inst);
1428 for (int i = 0; i < inst->numDestRegs(); i++) {
1430 DPRINTF(IEW,"Setting Destination Register %i (%s)\n",
1431 inst->renamedDestRegIdx(i)->index(),
1432 inst->renamedDestRegIdx(i)->className());
1433 scoreboard->setReg(inst->renamedDestRegIdx(i));
1437 producerInst[tid]++;
1438 consumerInst[tid]+= dependents;
1440 writebackCount[tid]++;
1445 template<class Impl>
1447 DefaultIEW<Impl>::tick()
1452 wroteToTimeBuffer = false;
1453 updatedQueues = false;
1459 // Free function units marked as being freed this cycle.
1460 fuPool->processFreeUnits();
1462 list<ThreadID>::iterator threads = activeThreads->begin();
1463 list<ThreadID>::iterator end = activeThreads->end();
1465 // Check stall and squash signals, dispatch any instructions.
1466 while (threads != end) {
1467 ThreadID tid = *threads++;
1469 DPRINTF(IEW,"Issue: Processing [tid:%i]\n",tid);
1471 checkSignalsAndUpdate(tid);
1475 if (exeStatus != Squashing) {
1480 // Have the instruction queue try to schedule any ready instructions.
1481 // (In actuality, this scheduling is for instructions that will
1482 // be executed next cycle.)
1483 instQueue.scheduleReadyInsts();
1485 // Also should advance its own time buffers if the stage ran.
1486 // Not the best place for it, but this works (hopefully).
1487 issueToExecQueue.advance();
1490 bool broadcast_free_entries = false;
1492 if (updatedQueues || exeStatus == Running || updateLSQNextCycle) {
1494 updateLSQNextCycle = false;
1496 broadcast_free_entries = true;
1499 // Writeback any stores using any leftover bandwidth.
1500 ldstQueue.writebackStores();
1502 // Check the committed load/store signals to see if there's a load
1503 // or store to commit. Also check if it's being told to execute a
1504 // nonspeculative instruction.
1505 // This is pretty inefficient...
1507 threads = activeThreads->begin();
1508 while (threads != end) {
1509 ThreadID tid = (*threads++);
1511 DPRINTF(IEW,"Processing [tid:%i]\n",tid);
1513 // Update structures based on instructions committed.
1514 if (fromCommit->commitInfo[tid].doneSeqNum != 0 &&
1515 !fromCommit->commitInfo[tid].squash &&
1516 !fromCommit->commitInfo[tid].robSquashing) {
1518 ldstQueue.commitStores(fromCommit->commitInfo[tid].doneSeqNum,tid);
1520 ldstQueue.commitLoads(fromCommit->commitInfo[tid].doneSeqNum,tid);
1522 updateLSQNextCycle = true;
1523 instQueue.commit(fromCommit->commitInfo[tid].doneSeqNum,tid);
1526 if (fromCommit->commitInfo[tid].nonSpecSeqNum != 0) {
1528 //DPRINTF(IEW,"NonspecInst from thread %i",tid);
1529 if (fromCommit->commitInfo[tid].strictlyOrdered) {
1530 instQueue.replayMemInst(
1531 fromCommit->commitInfo[tid].strictlyOrderedLoad);
1532 fromCommit->commitInfo[tid].strictlyOrderedLoad->setAtCommit();
1534 instQueue.scheduleNonSpec(
1535 fromCommit->commitInfo[tid].nonSpecSeqNum);
1539 if (broadcast_free_entries) {
1540 toFetch->iewInfo[tid].iqCount =
1541 instQueue.getCount(tid);
1542 toFetch->iewInfo[tid].ldstqCount =
1543 ldstQueue.getCount(tid);
1545 toRename->iewInfo[tid].usedIQ = true;
1546 toRename->iewInfo[tid].freeIQEntries =
1547 instQueue.numFreeEntries(tid);
1548 toRename->iewInfo[tid].usedLSQ = true;
1550 toRename->iewInfo[tid].freeLQEntries =
1551 ldstQueue.numFreeLoadEntries(tid);
1552 toRename->iewInfo[tid].freeSQEntries =
1553 ldstQueue.numFreeStoreEntries(tid);
1555 wroteToTimeBuffer = true;
1558 DPRINTF(IEW, "[tid:%i], Dispatch dispatched %i instructions.\n",
1559 tid, toRename->iewInfo[tid].dispatched);
1562 DPRINTF(IEW, "IQ has %i free entries (Can schedule: %i). "
1563 "LQ has %i free entries. SQ has %i free entries.\n",
1564 instQueue.numFreeEntries(), instQueue.hasReadyInsts(),
1565 ldstQueue.numFreeLoadEntries(), ldstQueue.numFreeStoreEntries());
1569 if (wroteToTimeBuffer) {
1570 DPRINTF(Activity, "Activity this cycle.\n");
1571 cpu->activityThisCycle();
1575 template <class Impl>
1577 DefaultIEW<Impl>::updateExeInstStats(const DynInstPtr& inst)
1579 ThreadID tid = inst->threadNumber;
1584 if (DTRACE(O3PipeView)) {
1585 inst->completeTick = curTick() - inst->fetchTick;
1590 // Control operations
1592 if (inst->isControl())
1593 iewExecutedBranches[tid]++;
1596 // Memory operations
1598 if (inst->isMemRef()) {
1599 iewExecutedRefs[tid]++;
1601 if (inst->isLoad()) {
1602 iewExecLoadInsts[tid]++;
1607 template <class Impl>
1609 DefaultIEW<Impl>::checkMisprediction(const DynInstPtr& inst)
1611 ThreadID tid = inst->threadNumber;
1613 if (!fetchRedirect[tid] ||
1614 !toCommit->squash[tid] ||
1615 toCommit->squashedSeqNum[tid] > inst->seqNum) {
1617 if (inst->mispredicted()) {
1618 fetchRedirect[tid] = true;
1620 DPRINTF(IEW, "Execute: Branch mispredict detected.\n");
1621 DPRINTF(IEW, "Predicted target was PC:%#x, NPC:%#x.\n",
1622 inst->predInstAddr(), inst->predNextInstAddr());
1623 DPRINTF(IEW, "Execute: Redirecting fetch to PC: %#x,"
1624 " NPC: %#x.\n", inst->nextInstAddr(),
1625 inst->nextInstAddr());
1626 // If incorrect, then signal the ROB that it must be squashed.
1627 squashDueToBranch(inst, tid);
1629 if (inst->readPredTaken()) {
1630 predictedTakenIncorrect++;
1632 predictedNotTakenIncorrect++;
1638 #endif//__CPU_O3_IEW_IMPL_IMPL_HH__