gem5/cpu/ozone/lw_back_end_impl.hh

#include "encumbered/cpu/full/op_class.hh"
#include "cpu/checker/cpu.hh"
#include "cpu/ozone/lw_back_end.hh"

template <class Impl>
void
LWBackEnd<Impl>::generateTrapEvent(Tick latency)
{
    DPRINTF(BE, "Generating trap event\n");

    TrapEvent *trap = new TrapEvent(this);

    trap->schedule(curTick + cpu->cycles(latency));

    thread->trapPending = true;
}

template <class Impl>
int
LWBackEnd<Impl>::wakeDependents(DynInstPtr &inst, bool memory_deps)
{
    assert(!inst->isSquashed());
    std::vector<DynInstPtr> &dependents = memory_deps ? inst->getMemDeps() :
        inst->getDependents();
    int num_outputs = dependents.size();

    DPRINTF(BE, "Waking instruction [sn:%lli] dependents in IQ\n", inst->seqNum);

    for (int i = 0; i < num_outputs; i++) {
        DynInstPtr dep_inst = dependents[i];
        if (!memory_deps) {
            dep_inst->markSrcRegReady();
        } else {
            if (!dep_inst->isSquashed())
                dep_inst->markMemInstReady(inst.get());
        }

        DPRINTF(BE, "Marking source reg ready [sn:%lli] in IQ\n", dep_inst->seqNum);

        if (dep_inst->readyToIssue() && dep_inst->isInROB() &&
            !dep_inst->isNonSpeculative() &&
            dep_inst->memDepReady() && !dep_inst->isMemBarrier() && !dep_inst->isWriteBarrier()) {
            DPRINTF(BE, "Adding instruction to exeList [sn:%lli]\n",
                    dep_inst->seqNum);
            exeList.push(dep_inst);
            if (dep_inst->iqItValid) {
                DPRINTF(BE, "Removing instruction from waiting list\n");
                waitingList.erase(dep_inst->iqIt);
                waitingInsts--;
                dep_inst->iqItValid = false;
                assert(waitingInsts >= 0);
            }
            if (dep_inst->isMemRef()) {
                removeWaitingMemOp(dep_inst);
                DPRINTF(BE, "Issued a waiting mem op [sn:%lli]\n",
                        dep_inst->seqNum);
            }
        }
    }
    return num_outputs;
}

template <class Impl>
void
LWBackEnd<Impl>::rescheduleMemInst(DynInstPtr &inst)
{
    replayList.push_front(inst);
}

template <class Impl>
LWBackEnd<Impl>::TrapEvent::TrapEvent(LWBackEnd<Impl> *_be)
    : Event(&mainEventQueue, CPU_Tick_Pri), be(_be)
{
    this->setFlags(Event::AutoDelete);
}

template <class Impl>
void
LWBackEnd<Impl>::TrapEvent::process()
{
    be->trapSquash = true;
}

template <class Impl>
const char *
LWBackEnd<Impl>::TrapEvent::description()
{
    return "Trap event";
}

template <class Impl>
void
LWBackEnd<Impl>::replayMemInst(DynInstPtr &inst)
{
    bool found_inst = false;
    while (!replayList.empty()) {
        exeList.push(replayList.front());
        if (replayList.front() == inst) {
            found_inst = true;
        }
        replayList.pop_front();
    }
    assert(found_inst);
}

template<class Impl>
LWBackEnd<Impl>::LdWritebackEvent::LdWritebackEvent(DynInstPtr &_inst,
                                                  LWBackEnd<Impl> *_be)
    : Event(&mainEventQueue), inst(_inst), be(_be), dcacheMiss(false)
{
    this->setFlags(Event::AutoDelete);
}

template<class Impl>
void
LWBackEnd<Impl>::LdWritebackEvent::process()
{
    DPRINTF(BE, "Load writeback event [sn:%lli]\n", inst->seqNum);
//    DPRINTF(Activity, "Activity: Ld Writeback event [sn:%lli]\n", inst->seqNum);

    //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum);

//    iewStage->wakeCPU();

    if (be->isSwitchedOut())
        return;

    if (dcacheMiss) {
        be->removeDcacheMiss(inst);
    }

    if (inst->isSquashed()) {
        inst = NULL;
        return;
    }

    if (!inst->isExecuted()) {
        inst->setExecuted();

        // Execute again to copy data to proper place.
        inst->completeAcc();
    }

    // Need to insert instruction into queue to commit
    be->instToCommit(inst);

    //wroteToTimeBuffer = true;
//    iewStage->activityThisCycle();

    inst = NULL;
}

template<class Impl>
const char *
LWBackEnd<Impl>::LdWritebackEvent::description()
{
    return "Load writeback event";
}


template <class Impl>
LWBackEnd<Impl>::DCacheCompletionEvent::DCacheCompletionEvent(LWBackEnd *_be)
    : Event(&mainEventQueue, CPU_Tick_Pri), be(_be)
{
}

template <class Impl>
void
LWBackEnd<Impl>::DCacheCompletionEvent::process()
{
}

template <class Impl>
const char *
LWBackEnd<Impl>::DCacheCompletionEvent::description()
{
    return "Cache completion event";
}

template <class Impl>
LWBackEnd<Impl>::LWBackEnd(Params *params)
    : d2i(5, 5), i2e(5, 5), e2c(5, 5), numInstsToWB(5, 5),
      trapSquash(false), xcSquash(false), cacheCompletionEvent(this),
      dcacheInterface(params->dcacheInterface), width(params->backEndWidth),
      exactFullStall(true)
{
    numROBEntries = params->numROBEntries;
    numInsts = 0;
    numDispatchEntries = 32;
    maxOutstandingMemOps = params->maxOutstandingMemOps;
    numWaitingMemOps = 0;
    waitingInsts = 0;
    switchedOut = false;
    switchPending = false;

//    IQ.setBE(this);
    LSQ.setBE(this);

    // Setup IQ and LSQ with their parameters here.
    instsToDispatch = d2i.getWire(-1);

    instsToExecute = i2e.getWire(-1);

//    IQ.setIssueExecQueue(&i2e);

    dispatchWidth = params->dispatchWidth ? params->dispatchWidth : width;
    issueWidth = params->issueWidth ? params->issueWidth : width;
    wbWidth = params->wbWidth ? params->wbWidth : width;
    commitWidth = params->commitWidth ? params->commitWidth : width;

    LSQ.init(params, params->LQEntries, params->SQEntries, 0);

    dispatchStatus = Running;
}

template <class Impl>
std::string
LWBackEnd<Impl>::name() const
{
    return cpu->name() + ".backend";
}

template <class Impl>
void
LWBackEnd<Impl>::regStats()
{
    using namespace Stats;
    rob_cap_events
        .init(cpu->number_of_threads)
        .name(name() + ".ROB:cap_events")
        .desc("number of cycles where ROB cap was active")
        .flags(total)
        ;

    rob_cap_inst_count
        .init(cpu->number_of_threads)
        .name(name() + ".ROB:cap_inst")
        .desc("number of instructions held up by ROB cap")
        .flags(total)
        ;

    iq_cap_events
        .init(cpu->number_of_threads)
        .name(name() +".IQ:cap_events" )
        .desc("number of cycles where IQ cap was active")
        .flags(total)
        ;

    iq_cap_inst_count
        .init(cpu->number_of_threads)
        .name(name() + ".IQ:cap_inst")
        .desc("number of instructions held up by IQ cap")
        .flags(total)
        ;


    exe_inst
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:count")
        .desc("number of insts issued")
        .flags(total)
        ;

    exe_swp
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:swp")
        .desc("number of swp insts issued")
        .flags(total)
        ;

    exe_nop
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:nop")
        .desc("number of nop insts issued")
        .flags(total)
        ;

    exe_refs
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:refs")
        .desc("number of memory reference insts issued")
        .flags(total)
        ;

    exe_loads
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:loads")
        .desc("number of load insts issued")
        .flags(total)
        ;

    exe_branches
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:branches")
        .desc("Number of branches issued")
        .flags(total)
        ;

    issued_ops
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:op_count")
        .desc("number of insts issued")
        .flags(total)
        ;

/*
    for (int i=0; i<Num_OpClasses; ++i) {
        stringstream subname;
        subname << opClassStrings[i] << "_delay";
        issue_delay_dist.subname(i, subname.str());
    }
*/
    //
    //  Other stats
    //
    lsq_forw_loads
        .init(cpu->number_of_threads)
        .name(name() + ".LSQ:forw_loads")
        .desc("number of loads forwarded via LSQ")
        .flags(total)
        ;

    inv_addr_loads
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:addr_loads")
        .desc("number of invalid-address loads")
        .flags(total)
        ;

    inv_addr_swpfs
        .init(cpu->number_of_threads)
        .name(name() + ".ISSUE:addr_swpfs")
        .desc("number of invalid-address SW prefetches")
        .flags(total)
        ;

    lsq_blocked_loads
        .init(cpu->number_of_threads)
        .name(name() + ".LSQ:blocked_loads")
        .desc("number of ready loads not issued due to memory disambiguation")
        .flags(total)
        ;

    lsqInversion
        .name(name() + ".ISSUE:lsq_invert")
        .desc("Number of times LSQ instruction issued early")
        ;

    n_issued_dist
        .init(issueWidth + 1)
        .name(name() + ".ISSUE:issued_per_cycle")
        .desc("Number of insts issued each cycle")
        .flags(total | pdf | dist)
        ;
    issue_delay_dist
        .init(Num_OpClasses,0,99,2)
        .name(name() + ".ISSUE:")
        .desc("cycles from operands ready to issue")
        .flags(pdf | cdf)
        ;

    queue_res_dist
        .init(Num_OpClasses, 0, 99, 2)
        .name(name() + ".IQ:residence:")
        .desc("cycles from dispatch to issue")
        .flags(total | pdf | cdf )
        ;
    for (int i = 0; i < Num_OpClasses; ++i) {
        queue_res_dist.subname(i, opClassStrings[i]);
    }

    writeback_count
        .init(cpu->number_of_threads)
        .name(name() + ".WB:count")
        .desc("cumulative count of insts written-back")
        .flags(total)
        ;

    producer_inst
        .init(cpu->number_of_threads)
        .name(name() + ".WB:producers")
        .desc("num instructions producing a value")
        .flags(total)
        ;

    consumer_inst
        .init(cpu->number_of_threads)
        .name(name() + ".WB:consumers")
        .desc("num instructions consuming a value")
        .flags(total)
        ;

    wb_penalized
        .init(cpu->number_of_threads)
        .name(name() + ".WB:penalized")
        .desc("number of instrctions required to write to 'other' IQ")
        .flags(total)
        ;


    wb_penalized_rate
        .name(name() + ".WB:penalized_rate")
        .desc ("fraction of instructions written-back that wrote to 'other' IQ")
        .flags(total)
        ;

    wb_penalized_rate = wb_penalized / writeback_count;

    wb_fanout
        .name(name() + ".WB:fanout")
        .desc("average fanout of values written-back")
        .flags(total)
        ;

    wb_fanout = producer_inst / consumer_inst;

    wb_rate
        .name(name() + ".WB:rate")
        .desc("insts written-back per cycle")
        .flags(total)
        ;
    wb_rate = writeback_count / cpu->numCycles;

    stat_com_inst
        .init(cpu->number_of_threads)
        .name(name() + ".COM:count")
        .desc("Number of instructions committed")
        .flags(total)
        ;

    stat_com_swp
        .init(cpu->number_of_threads)
        .name(name() + ".COM:swp_count")
        .desc("Number of s/w prefetches committed")
        .flags(total)
        ;

    stat_com_refs
        .init(cpu->number_of_threads)
        .name(name() +  ".COM:refs")
        .desc("Number of memory references committed")
        .flags(total)
        ;

    stat_com_loads
        .init(cpu->number_of_threads)
        .name(name() +  ".COM:loads")
        .desc("Number of loads committed")
        .flags(total)
        ;

    stat_com_membars
        .init(cpu->number_of_threads)
        .name(name() +  ".COM:membars")
        .desc("Number of memory barriers committed")
        .flags(total)
        ;

    stat_com_branches
        .init(cpu->number_of_threads)
        .name(name() + ".COM:branches")
        .desc("Number of branches committed")
        .flags(total)
        ;
    n_committed_dist
        .init(0,commitWidth,1)
        .name(name() + ".COM:committed_per_cycle")
        .desc("Number of insts commited each cycle")
        .flags(pdf)
        ;

    //
    //  Commit-Eligible instructions...
    //
    //  -> The number of instructions eligible to commit in those
    //  cycles where we reached our commit BW limit (less the number
    //  actually committed)
    //
    //  -> The average value is computed over ALL CYCLES... not just
    //  the BW limited cycles
    //
    //  -> The standard deviation is computed only over cycles where
    //  we reached the BW limit
    //
    commit_eligible
        .init(cpu->number_of_threads)
        .name(name() + ".COM:bw_limited")
        .desc("number of insts not committed due to BW limits")
        .flags(total)
        ;

    commit_eligible_samples
        .name(name() + ".COM:bw_lim_events")
        .desc("number cycles where commit BW limit reached")
        ;

    squashedInsts
        .init(cpu->number_of_threads)
        .name(name() + ".COM:squashed_insts")
        .desc("Number of instructions removed from inst list")
        ;

    ROBSquashedInsts
        .init(cpu->number_of_threads)
        .name(name() + ".COM:rob_squashed_insts")
        .desc("Number of instructions removed from inst list when they reached the head of the ROB")
        ;

    ROB_fcount
        .name(name() + ".ROB:full_count")
        .desc("number of cycles where ROB was full")
        ;

    ROB_count
        .init(cpu->number_of_threads)
        .name(name() + ".ROB:occupancy")
        .desc(name() + ".ROB occupancy (cumulative)")
        .flags(total)
        ;

    ROB_full_rate
        .name(name() + ".ROB:full_rate")
        .desc("ROB full per cycle")
        ;
    ROB_full_rate = ROB_fcount / cpu->numCycles;

    ROB_occ_rate
        .name(name() + ".ROB:occ_rate")
        .desc("ROB occupancy rate")
        .flags(total)
        ;
    ROB_occ_rate = ROB_count / cpu->numCycles;

    ROB_occ_dist
        .init(cpu->number_of_threads,0,numROBEntries,2)
        .name(name() + ".ROB:occ_dist")
        .desc("ROB Occupancy per cycle")
        .flags(total | cdf)
        ;

//    IQ.regStats();
}

template <class Impl>
void
LWBackEnd<Impl>::setCPU(FullCPU *cpu_ptr)
{
    cpu = cpu_ptr;
    LSQ.setCPU(cpu_ptr);
    checker = cpu->checker;
}

template <class Impl>
void
LWBackEnd<Impl>::setCommBuffer(TimeBuffer<CommStruct> *_comm)
{
    comm = _comm;
    toIEW = comm->getWire(0);
    fromCommit = comm->getWire(-1);
}

#if FULL_SYSTEM
template <class Impl>
void
LWBackEnd<Impl>::checkInterrupts()
{
    if (cpu->checkInterrupts &&
        cpu->check_interrupts() &&
        !cpu->inPalMode(thread->readPC()) &&
        !trapSquash &&
        !xcSquash) {
        frontEnd->interruptPending = true;
        if (robEmpty() && !LSQ.hasStoresToWB()) {
            // Will need to squash all instructions currently in flight and have
            // the interrupt handler restart at the last non-committed inst.
            // Most of that can be handled through the trap() function.  The
            // processInterrupts() function really just checks for interrupts
            // and then calls trap() if there is an interrupt present.

            // Not sure which thread should be the one to interrupt.  For now
            // always do thread 0.
            assert(!thread->inSyscall);
            thread->inSyscall = true;

            // CPU will handle implementation of the interrupt.
            cpu->processInterrupts();

            // Now squash or record that I need to squash this cycle.
            commitStatus = TrapPending;

            // Exit state update mode to avoid accidental updating.
            thread->inSyscall = false;

            // Generate trap squash event.
            generateTrapEvent();

            DPRINTF(BE, "Interrupt detected.\n");
        } else {
            DPRINTF(BE, "Interrupt must wait for ROB to drain.\n");
        }
    }
}

template <class Impl>
void
LWBackEnd<Impl>::handleFault(Fault &fault, Tick latency)
{
    DPRINTF(BE, "Handling fault!\n");

    assert(!thread->inSyscall);

    thread->inSyscall = true;

    // Consider holding onto the trap and waiting until the trap event
    // happens for this to be executed.
    fault->invoke(thread->getXCProxy());

    // Exit state update mode to avoid accidental updating.
    thread->inSyscall = false;

    commitStatus = TrapPending;

    // Generate trap squash event.
    generateTrapEvent(latency);
}
#endif

template <class Impl>
void
LWBackEnd<Impl>::tick()
{
    DPRINTF(BE, "Ticking back end\n");

    if (switchPending && robEmpty() && !LSQ.hasStoresToWB()) {
        cpu->signalSwitched();
        return;
    }

    ROB_count[0]+= numInsts;

    wbCycle = 0;

    // Read in any done instruction information and update the IQ or LSQ.
    updateStructures();

#if FULL_SYSTEM
    checkInterrupts();

    if (trapSquash) {
        assert(!xcSquash);
        squashFromTrap();
    } else if (xcSquash) {
        squashFromXC();
    } /*else if (fetchHasFault && robEmpty() && frontEnd->isEmpty() && !LSQ.hasStoresToWB()) {
        DPRINTF(BE, "ROB and front end empty, handling fetch fault\n");
        Fault fetch_fault = frontEnd->getFault();
        if (fetch_fault == NoFault) {
            DPRINTF(BE, "Fetch no longer has a fault, cancelling out.\n");
            fetchHasFault = false;
        } else {
            handleFault(fetch_fault);
            fetchHasFault = false;
        }
        }*/
#endif

    if (dispatchStatus != Blocked) {
        dispatchInsts();
    } else {
        checkDispatchStatus();
    }

    if (commitStatus != TrapPending) {
        executeInsts();

        commitInsts();
    }

    LSQ.writebackStores();

    DPRINTF(BE, "Waiting insts: %i, mem ops: %i, ROB entries in use: %i, "
            "LSQ loads: %i, LSQ stores: %i\n",
            waitingInsts, numWaitingMemOps, numInsts,
            LSQ.numLoads(), LSQ.numStores());

#ifdef DEBUG
    assert(numInsts == instList.size());
    assert(waitingInsts == waitingList.size());
    assert(numWaitingMemOps == waitingMemOps.size());
    assert(!switchedOut);
#endif
}

template <class Impl>
void
LWBackEnd<Impl>::updateStructures()
{
    if (fromCommit->doneSeqNum) {
        LSQ.commitLoads(fromCommit->doneSeqNum);
        LSQ.commitStores(fromCommit->doneSeqNum);
    }

    if (fromCommit->nonSpecSeqNum) {
        if (fromCommit->uncached) {
//            LSQ.executeLoad(fromCommit->lqIdx);
        } else {
//            IQ.scheduleNonSpec(
//                fromCommit->nonSpecSeqNum);
        }
    }
}

template <class Impl>
void
LWBackEnd<Impl>::addToLSQ(DynInstPtr &inst)
{
    // Do anything LSQ specific here?
    LSQ.insert(inst);
}

template <class Impl>
void
LWBackEnd<Impl>::dispatchInsts()
{
    DPRINTF(BE, "Trying to dispatch instructions.\n");

    while (numInsts < numROBEntries &&
           numWaitingMemOps < maxOutstandingMemOps) {
        // Get instruction from front of time buffer
        DynInstPtr inst = frontEnd->getInst();
        if (!inst) {
            break;
        } else if (inst->isSquashed()) {
            continue;
        }

        ++numInsts;
        instList.push_front(inst);

        inst->setInROB();

        DPRINTF(BE, "Dispatching instruction [sn:%lli] PC:%#x\n",
                inst->seqNum, inst->readPC());

        for (int i = 0; i < inst->numDestRegs(); ++i)
            renameTable[inst->destRegIdx(i)] = inst;

        if (inst->isMemBarrier() || inst->isWriteBarrier()) {
            if (memBarrier) {
                DPRINTF(BE, "Instruction [sn:%lli] is waiting on "
                        "barrier [sn:%lli].\n",
                        inst->seqNum, memBarrier->seqNum);
                memBarrier->addMemDependent(inst);
                inst->addSrcMemInst(memBarrier);
            }
            memBarrier = inst;
            inst->setCanCommit();
        } else if (inst->readyToIssue() && !inst->isNonSpeculative()) {
            if (inst->isMemRef()) {

                LSQ.insert(inst);
                if (memBarrier) {
                    DPRINTF(BE, "Instruction [sn:%lli] is waiting on "
                            "barrier [sn:%lli].\n",
                            inst->seqNum, memBarrier->seqNum);
                    memBarrier->addMemDependent(inst);
                    inst->addSrcMemInst(memBarrier);
                    addWaitingMemOp(inst);

                    waitingList.push_front(inst);
                    inst->iqIt = waitingList.begin();
                    inst->iqItValid = true;
                    waitingInsts++;
                } else {
                    DPRINTF(BE, "Instruction [sn:%lli] ready, addding to exeList.\n",
                            inst->seqNum);
                    exeList.push(inst);
                }
            } else if (inst->isNop()) {
                DPRINTF(BE, "Nop encountered [sn:%lli], skipping exeList.\n",
                        inst->seqNum);
                inst->setIssued();
                inst->setExecuted();
                inst->setCanCommit();
            } else {
                DPRINTF(BE, "Instruction [sn:%lli] ready, addding to exeList.\n",
                        inst->seqNum);
                exeList.push(inst);
            }
        } else {
            if (inst->isNonSpeculative()) {
                inst->setCanCommit();
                DPRINTF(BE, "Adding non speculative instruction\n");
            }

            if (inst->isMemRef()) {
                addWaitingMemOp(inst);
                LSQ.insert(inst);
                if (memBarrier) {
                    memBarrier->addMemDependent(inst);
                    inst->addSrcMemInst(memBarrier);

                    DPRINTF(BE, "Instruction [sn:%lli] is waiting on "
                            "barrier [sn:%lli].\n",
                            inst->seqNum, memBarrier->seqNum);
                }
            }

            DPRINTF(BE, "Instruction [sn:%lli] not ready, addding to "
                    "waitingList.\n",
                    inst->seqNum);
            waitingList.push_front(inst);
            inst->iqIt = waitingList.begin();
            inst->iqItValid = true;
            waitingInsts++;
        }
    }

    // Check if IQ or LSQ is full.  If so we'll need to break and stop
    // removing instructions.  Also update the number of insts to remove
    // from the queue.  Check here if we don't care about exact stall
    // conditions.
/*
    bool stall = false;
    if (IQ.isFull()) {
        DPRINTF(BE, "IQ is full!\n");
        stall = true;
    } else if (LSQ.isFull()) {
        DPRINTF(BE, "LSQ is full!\n");
        stall = true;
    } else if (isFull()) {
        DPRINTF(BE, "ROB is full!\n");
        stall = true;
        ROB_fcount++;
    }
    if (stall) {
        d2i.advance();
        dispatchStall();
        return;
    }
*/
}

template <class Impl>
void
LWBackEnd<Impl>::dispatchStall()
{
    dispatchStatus = Blocked;
    if (!cpu->decoupledFrontEnd) {
        // Tell front end to stall here through a timebuffer, or just tell
        // it directly.
    }
}

template <class Impl>
void
LWBackEnd<Impl>::checkDispatchStatus()
{
    DPRINTF(BE, "Checking dispatch status\n");
    assert(dispatchStatus == Blocked);
    if (!LSQ.isFull() && !isFull()) {
        DPRINTF(BE, "Dispatch no longer blocked\n");
        dispatchStatus = Running;
        dispatchInsts();
    }
}

template <class Impl>
void
LWBackEnd<Impl>::executeInsts()
{
    DPRINTF(BE, "Trying to execute instructions\n");

    int num_executed = 0;
    while (!exeList.empty() && num_executed < issueWidth) {
        DynInstPtr inst = exeList.top();

        DPRINTF(BE, "Executing inst [sn:%lli] PC: %#x\n",
                inst->seqNum, inst->readPC());

        // Check if the instruction is squashed; if so then skip it
        // and don't count it towards the FU usage.
        if (inst->isSquashed()) {
            DPRINTF(BE, "Execute: Instruction was squashed.\n");

            // Not sure how to handle this plus the method of sending # of
            // instructions to use.  Probably will just have to count it
            // towards the bandwidth usage, but not the FU usage.
            ++num_executed;

            // Consider this instruction executed so that commit can go
            // ahead and retire the instruction.
            inst->setExecuted();

            // Not sure if I should set this here or just let commit try to
            // commit any squashed instructions.  I like the latter a bit more.
            inst->setCanCommit();

//            ++iewExecSquashedInsts;
            exeList.pop();

            continue;
        }

        Fault fault = NoFault;

        // Execute instruction.
        // Note that if the instruction faults, it will be handled
        // at the commit stage.
        if (inst->isMemRef() &&
            (!inst->isDataPrefetch() && !inst->isInstPrefetch())) {
            if (dcacheInterface->isBlocked()) {
                // Should I move the instruction aside?
                DPRINTF(BE, "Execute: dcache is blocked\n");
                break;
            }
            DPRINTF(BE, "Execute: Initiating access for memory "
                    "reference.\n");

            if (inst->isLoad()) {
                LSQ.executeLoad(inst);
            } else if (inst->isStore()) {
                LSQ.executeStore(inst);
                if (inst->req && !(inst->req->flags & LOCKED)) {
                    inst->setExecuted();

                    instToCommit(inst);
                }
            } else {
                panic("Unknown mem type!");
            }
        } else {
            inst->execute();

            inst->setExecuted();

            instToCommit(inst);
        }

        updateExeInstStats(inst);

        ++funcExeInst;
        ++num_executed;

        exeList.pop();

        if (inst->mispredicted()) {
            squashDueToBranch(inst);
            break;
        } else if (LSQ.violation()) {
            // Get the DynInst that caused the violation.  Note that this
            // clears the violation signal.
            DynInstPtr violator;
            violator = LSQ.getMemDepViolator();

            DPRINTF(BE, "LDSTQ detected a violation.  Violator PC: "
                    "%#x, inst PC: %#x.  Addr is: %#x.\n",
                    violator->readPC(), inst->readPC(), inst->physEffAddr);

            // Squash.
            squashDueToMemViolation(inst);
        }
    }

    issued_ops[0]+= num_executed;
    n_issued_dist[num_executed]++;
}

template<class Impl>
void
LWBackEnd<Impl>::instToCommit(DynInstPtr &inst)
{

    DPRINTF(BE, "Sending instructions to commit [sn:%lli] PC %#x.\n",
            inst->seqNum, inst->readPC());

    if (!inst->isSquashed()) {
        DPRINTF(BE, "Writing back instruction [sn:%lli] PC %#x.\n",
                inst->seqNum, inst->readPC());

        inst->setCanCommit();

        if (inst->isExecuted()) {
            inst->setResultReady();
            int dependents = wakeDependents(inst);
            if (dependents) {
                producer_inst[0]++;
                consumer_inst[0]+= dependents;
            }
        }
    }

    writeback_count[0]++;
}

template <class Impl>
void
LWBackEnd<Impl>::writebackInsts()
{
    int wb_width = wbWidth;
    // Using this method I'm not quite sure how to prevent an
    // instruction from waking its own dependents multiple times,
    // without the guarantee that commit always has enough bandwidth
    // to accept all instructions being written back.  This guarantee
    // might not be too unrealistic.
    InstListIt wb_inst_it = writeback.begin();
    InstListIt wb_end_it = writeback.end();
    int inst_num = 0;
    int consumer_insts = 0;

    for (; inst_num < wb_width &&
             wb_inst_it != wb_end_it; inst_num++) {
        DynInstPtr inst = (*wb_inst_it);

        // Some instructions will be sent to commit without having
        // executed because they need commit to handle them.
        // E.g. Uncached loads have not actually executed when they
        // are first sent to commit.  Instead commit must tell the LSQ
        // when it's ready to execute the uncached load.
        if (!inst->isSquashed()) {
            DPRINTF(BE, "Writing back instruction [sn:%lli] PC %#x.\n",
                    inst->seqNum, inst->readPC());

            inst->setCanCommit();
            inst->setResultReady();

            if (inst->isExecuted()) {
                int dependents = wakeDependents(inst);
                if (dependents) {
                    producer_inst[0]++;
                    consumer_insts+= dependents;
                }
            }
        }

        writeback.erase(wb_inst_it++);
    }
    LSQ.writebackStores();
    consumer_inst[0]+= consumer_insts;
    writeback_count[0]+= inst_num;
}

template <class Impl>
bool
LWBackEnd<Impl>::commitInst(int inst_num)
{
    // Read instruction from the head of the ROB
    DynInstPtr inst = instList.back();

    // Make sure instruction is valid
    assert(inst);

    if (!inst->readyToCommit())
        return false;

    DPRINTF(BE, "Trying to commit instruction [sn:%lli] PC:%#x\n",
            inst->seqNum, inst->readPC());

    thread->setPC(inst->readPC());
    thread->setNextPC(inst->readNextPC());
    inst->reachedCommit = true;

    // If the instruction is not executed yet, then it is a non-speculative
    // or store inst.  Signal backwards that it should be executed.
    if (!inst->isExecuted()) {
        if (inst->isNonSpeculative() ||
            inst->isMemBarrier() ||
            inst->isWriteBarrier()) {
#if !FULL_SYSTEM
            // Hack to make sure syscalls aren't executed until all stores
            // write back their data.  This direct communication shouldn't
            // be used for anything other than this.
            if (inst_num > 0 || LSQ.hasStoresToWB())
#else
            if ((inst->isMemBarrier() || inst->isWriteBarrier() ||
                    inst->isQuiesce()) &&
                LSQ.hasStoresToWB())
#endif
            {
                DPRINTF(BE, "Waiting for all stores to writeback.\n");
                return false;
            }

            DPRINTF(BE, "Encountered a store or non-speculative "
                    "instruction at the head of the ROB, PC %#x.\n",
                    inst->readPC());

            if (inst->isMemBarrier() || inst->isWriteBarrier()) {
                DPRINTF(BE, "Waking dependents on barrier [sn:%lli]\n",
                        inst->seqNum);
                assert(memBarrier);
                wakeDependents(inst, true);
                if (memBarrier == inst)
                    memBarrier = NULL;
                inst->clearMemDependents();
            }

            // Send back the non-speculative instruction's sequence number.
            if (inst->iqItValid) {
                DPRINTF(BE, "Removing instruction from waiting list\n");
                waitingList.erase(inst->iqIt);
                inst->iqItValid = false;
                waitingInsts--;
                assert(waitingInsts >= 0);
                if (inst->isStore())
                    removeWaitingMemOp(inst);
            }

            exeList.push(inst);

            // Change the instruction so it won't try to commit again until
            // it is executed.
            inst->clearCanCommit();

//            ++commitNonSpecStalls;

            return false;
        } else if (inst->isLoad()) {
            DPRINTF(BE, "[sn:%lli]: Uncached load, PC %#x.\n",
                    inst->seqNum, inst->readPC());

            // Send back the non-speculative instruction's sequence
            // number.  Maybe just tell the lsq to re-execute the load.

            // Send back the non-speculative instruction's sequence number.
            if (inst->iqItValid) {
                DPRINTF(BE, "Removing instruction from waiting list\n");
                waitingList.erase(inst->iqIt);
                inst->iqItValid = false;
                waitingInsts--;
                assert(waitingInsts >= 0);
                removeWaitingMemOp(inst);
            }
            replayMemInst(inst);

            inst->clearCanCommit();

            return false;
        } else {
            panic("Trying to commit un-executed instruction "
                  "of unknown type!\n");
        }
    }

    // Not handled for now.
    assert(!inst->isThreadSync());
    assert(inst->memDepReady());
    // Stores will mark themselves as totally completed as they need
    // to wait to writeback to memory.  @todo: Hack...attempt to fix
    // having the checker be forced to wait until a store completes in
    // order to check all of the instructions.  If the store at the
    // head of the check list misses, but a later store hits, then
    // loads in the checker may see the younger store values instead
    // of the store they should see.  Either the checker needs its own
    // memory (annoying to update), its own store buffer (how to tell
    // which value is correct?), or something else...
    if (!inst->isStore()) {
        inst->setCompleted();
    }
    // Check if the instruction caused a fault.  If so, trap.
    Fault inst_fault = inst->getFault();

    // Use checker prior to updating anything due to traps or PC
    // based events.
    if (checker) {
        checker->tick(inst);
    }

    if (inst_fault != NoFault) {
        DPRINTF(BE, "Inst [sn:%lli] PC %#x has a fault\n",
                inst->seqNum, inst->readPC());

        // Instruction is completed as it has a fault.
        inst->setCompleted();

        if (LSQ.hasStoresToWB()) {
            DPRINTF(BE, "Stores still in flight, will wait until drained.\n");
            return false;
        } else if (inst_num != 0) {
            DPRINTF(BE, "Will wait until instruction is head of commit group.\n");
            return false;
        } else if (checker && inst->isStore()) {
            checker->tick(inst);
        }

        thread->setInst(
            static_cast<TheISA::MachInst>(inst->staticInst->machInst));
#if FULL_SYSTEM
        handleFault(inst_fault);
        return false;
#else // !FULL_SYSTEM
        panic("fault (%d) detected @ PC %08p", inst_fault,
              inst->PC);
#endif // FULL_SYSTEM
    }

    int freed_regs = 0;

    for (int i = 0; i < inst->numDestRegs(); ++i) {
        DPRINTF(BE, "Commit rename map setting reg %i to [sn:%lli]\n",
                (int)inst->destRegIdx(i), inst->seqNum);
        thread->renameTable[inst->destRegIdx(i)] = inst;
        ++freed_regs;
    }

    if (inst->traceData) {
        inst->traceData->setFetchSeq(inst->seqNum);
        inst->traceData->setCPSeq(thread->numInst);
        inst->traceData->finalize();
        inst->traceData = NULL;
    }

    inst->clearDependents();

    frontEnd->addFreeRegs(freed_regs);

    instList.pop_back();

    --numInsts;
    ++thread->funcExeInst;
    // Maybe move this to where the fault is handled; if the fault is handled,
    // don't try to set this myself as the fault will set it.  If not, then
    // I set thread->PC = thread->nextPC and thread->nextPC = thread->nextPC + 4.
    thread->setPC(thread->readNextPC());
    thread->setNextPC(thread->readNextPC() + sizeof(TheISA::MachInst));
    updateComInstStats(inst);

    // Write the done sequence number here.
//    LSQ.commitLoads(inst->seqNum);
    toIEW->doneSeqNum = inst->seqNum;
    lastCommitCycle = curTick;

#if FULL_SYSTEM
    int count = 0;
    Addr oldpc;
    do {
        if (count == 0)
            assert(!thread->inSyscall && !thread->trapPending);
        oldpc = thread->readPC();
        cpu->system->pcEventQueue.service(
            thread->getXCProxy());
        count++;
    } while (oldpc != thread->readPC());
    if (count > 1) {
        DPRINTF(BE, "PC skip function event, stopping commit\n");
        xcSquash = true;
        return false;
    }
#endif
    return true;
}

template <class Impl>
void
LWBackEnd<Impl>::commitInsts()
{
    // Not sure this should be a loop or not.
    int inst_num = 0;
    while (!instList.empty() && inst_num < commitWidth) {
        if (instList.back()->isSquashed()) {
            instList.back()->clearDependents();
            instList.pop_back();
            --numInsts;
            ROBSquashedInsts[instList.back()->threadNumber]++;
            continue;
        }

        if (!commitInst(inst_num++)) {
            DPRINTF(BE, "Can't commit, Instruction [sn:%lli] PC "
                    "%#x is head of ROB and not ready\n",
                    instList.back()->seqNum, instList.back()->readPC());
            --inst_num;
            break;
        }
    }
    n_committed_dist.sample(inst_num);
}

template <class Impl>
void
LWBackEnd<Impl>::squash(const InstSeqNum &sn)
{
    LSQ.squash(sn);

    int freed_regs = 0;
    InstListIt waiting_list_end = waitingList.end();
    InstListIt insts_it = waitingList.begin();

    while (insts_it != waiting_list_end && (*insts_it)->seqNum > sn)
    {
        if ((*insts_it)->isSquashed()) {
            ++insts_it;
            continue;
        }
        DPRINTF(BE, "Squashing instruction on waitingList PC %#x, [sn:%lli].\n",
                (*insts_it)->readPC(),
                (*insts_it)->seqNum);

        if ((*insts_it)->isMemRef()) {
            DPRINTF(BE, "Squashing a waiting mem op [sn:%lli]\n",
                    (*insts_it)->seqNum);
            removeWaitingMemOp((*insts_it));
        }

        waitingList.erase(insts_it++);
        waitingInsts--;
    }
    assert(waitingInsts >= 0);

    insts_it = instList.begin();

    while (!instList.empty() && (*insts_it)->seqNum > sn)
    {
        if ((*insts_it)->isSquashed()) {
            ++insts_it;
            continue;
        }
        DPRINTF(BE, "Squashing instruction on inst list PC %#x, [sn:%lli].\n",
                (*insts_it)->readPC(),
                (*insts_it)->seqNum);

        // Mark the instruction as squashed, and ready to commit so that
        // it can drain out of the pipeline.
        (*insts_it)->setSquashed();

        (*insts_it)->setCanCommit();

        (*insts_it)->removeInROB();

        for (int i = 0; i < (*insts_it)->numDestRegs(); ++i) {
            DynInstPtr prev_dest = (*insts_it)->getPrevDestInst(i);
            DPRINTF(BE, "Commit rename map setting reg %i to [sn:%lli]\n",
                    (int)(*insts_it)->destRegIdx(i), prev_dest->seqNum);
            renameTable[(*insts_it)->destRegIdx(i)] = prev_dest;
            ++freed_regs;
        }

        (*insts_it)->clearDependents();

        squashedInsts[(*insts_it)->threadNumber]++;

        instList.erase(insts_it++);
        --numInsts;
    }

    insts_it = waitingList.begin();
    while (!waitingList.empty() && insts_it != waitingList.end()) {
        if ((*insts_it)->seqNum < sn) {
            ++insts_it;
            continue;
        }
        assert((*insts_it)->isSquashed());

        waitingList.erase(insts_it++);
        waitingInsts--;
    }

    while (memBarrier && memBarrier->seqNum > sn) {
        DPRINTF(BE, "[sn:%lli] Memory barrier squashed (or previously squashed)\n", memBarrier->seqNum);
        memBarrier->clearMemDependents();
        if (memBarrier->memDepReady()) {
            DPRINTF(BE, "No previous barrier\n");
            memBarrier = NULL;
        } else {
            std::list<DynInstPtr> &srcs = memBarrier->getMemSrcs();
            memBarrier = srcs.front();
            srcs.pop_front();
            assert(srcs.empty());
            DPRINTF(BE, "Previous barrier: [sn:%lli]\n",
                    memBarrier->seqNum);
        }
    }

    frontEnd->addFreeRegs(freed_regs);
}

template <class Impl>
void
LWBackEnd<Impl>::squashFromXC()
{
    InstSeqNum squashed_inst = robEmpty() ? 0 : instList.back()->seqNum - 1;
    squash(squashed_inst);
    frontEnd->squash(squashed_inst, thread->readPC(),
                     false, false);
    frontEnd->interruptPending = false;

    thread->trapPending = false;
    thread->inSyscall = false;
    xcSquash = false;
    commitStatus = Running;
}

template <class Impl>
void
LWBackEnd<Impl>::squashFromTrap()
{
    InstSeqNum squashed_inst = robEmpty() ? 0 : instList.back()->seqNum - 1;
    squash(squashed_inst);
    frontEnd->squash(squashed_inst, thread->readPC(),
                     false, false);
    frontEnd->interruptPending = false;

    thread->trapPending = false;
    thread->inSyscall = false;
    trapSquash = false;
    commitStatus = Running;
}

template <class Impl>
void
LWBackEnd<Impl>::squashDueToBranch(DynInstPtr &inst)
{
    // Update the branch predictor state I guess
    DPRINTF(BE, "Squashing due to branch [sn:%lli], will restart at PC %#x\n",
            inst->seqNum, inst->readNextPC());
    squash(inst->seqNum);
    frontEnd->squash(inst->seqNum, inst->readNextPC(),
                     true, inst->mispredicted());
}

template <class Impl>
void
LWBackEnd<Impl>::squashDueToMemViolation(DynInstPtr &inst)
{
    // Update the branch predictor state I guess
    DPRINTF(BE, "Squashing due to violation [sn:%lli], will restart at PC %#x\n",
            inst->seqNum, inst->readNextPC());
    squash(inst->seqNum);
    frontEnd->squash(inst->seqNum, inst->readNextPC(),
                     false, inst->mispredicted());
}

template <class Impl>
void
LWBackEnd<Impl>::squashDueToMemBlocked(DynInstPtr &inst)
{
    DPRINTF(IEW, "Memory blocked, squashing load and younger insts, "
            "PC: %#x [sn:%i].\n", inst->readPC(), inst->seqNum);

    squash(inst->seqNum - 1);
    frontEnd->squash(inst->seqNum - 1, inst->readPC());
}

template <class Impl>
void
LWBackEnd<Impl>::fetchFault(Fault &fault)
{
    faultFromFetch = fault;
    fetchHasFault = true;
}

template <class Impl>
void
LWBackEnd<Impl>::switchOut()
{
    switchPending = true;
}

template <class Impl>
void
LWBackEnd<Impl>::doSwitchOut()
{
    switchedOut = true;
    switchPending = false;
    // Need to get rid of all committed, non-speculative state and write it
    // to memory/XC.  In this case this is stores that have committed and not
    // yet written back.
    assert(robEmpty());
    assert(!LSQ.hasStoresToWB());

    LSQ.switchOut();

    squash(0);
}

template <class Impl>
void
LWBackEnd<Impl>::takeOverFrom(ExecContext *old_xc)
{
    switchedOut = false;
    xcSquash = false;
    trapSquash = false;

    numInsts = 0;
    numWaitingMemOps = 0;
    waitingMemOps.clear();
    waitingInsts = 0;
    switchedOut = false;
    dispatchStatus = Running;
    commitStatus = Running;
    LSQ.takeOverFrom(old_xc);
}

template <class Impl>
void
LWBackEnd<Impl>::updateExeInstStats(DynInstPtr &inst)
{
    int thread_number = inst->threadNumber;

    //
    //  Pick off the software prefetches
    //
#ifdef TARGET_ALPHA
    if (inst->isDataPrefetch())
        exe_swp[thread_number]++;
    else
        exe_inst[thread_number]++;
#else
    exe_inst[thread_number]++;
#endif

    //
    //  Control operations
    //
    if (inst->isControl())
        exe_branches[thread_number]++;

    //
    //  Memory operations
    //
    if (inst->isMemRef()) {
        exe_refs[thread_number]++;

        if (inst->isLoad())
            exe_loads[thread_number]++;
    }
}

template <class Impl>
void
LWBackEnd<Impl>::updateComInstStats(DynInstPtr &inst)
{
    unsigned tid = inst->threadNumber;

    // keep an instruction count
    thread->numInst++;
    thread->numInsts++;

    cpu->numInst++;
    //
    //  Pick off the software prefetches
    //
#ifdef TARGET_ALPHA
    if (inst->isDataPrefetch()) {
        stat_com_swp[tid]++;
    } else {
        stat_com_inst[tid]++;
    }
#else
    stat_com_inst[tid]++;
#endif

    //
    //  Control Instructions
    //
    if (inst->isControl())
        stat_com_branches[tid]++;

    //
    //  Memory references
    //
    if (inst->isMemRef()) {
        stat_com_refs[tid]++;

        if (inst->isLoad()) {
            stat_com_loads[tid]++;
        }
    }

    if (inst->isMemBarrier()) {
        stat_com_membars[tid]++;
    }
}

template <class Impl>
void
LWBackEnd<Impl>::dumpInsts()
{
    int num = 0;
    int valid_num = 0;

    InstListIt inst_list_it = --(instList.end());

    cprintf("ExeList size: %i\n", exeList.size());

    cprintf("Inst list size: %i\n", instList.size());

    while (inst_list_it != instList.end())
    {
        cprintf("Instruction:%i\n",
                num);
        if (!(*inst_list_it)->isSquashed()) {
            if (!(*inst_list_it)->isIssued()) {
                ++valid_num;
                cprintf("Count:%i\n", valid_num);
            } else if ((*inst_list_it)->isMemRef() &&
                       !(*inst_list_it)->memOpDone) {
                // Loads that have not been marked as executed still count
                // towards the total instructions.
                ++valid_num;
                cprintf("Count:%i\n", valid_num);
            }
        }

        cprintf("PC:%#x\n[sn:%lli]\n[tid:%i]\n"
                "Issued:%i\nSquashed:%i\n",
                (*inst_list_it)->readPC(),
                (*inst_list_it)->seqNum,
                (*inst_list_it)->threadNumber,
                (*inst_list_it)->isIssued(),
                (*inst_list_it)->isSquashed());

        if ((*inst_list_it)->isMemRef()) {
            cprintf("MemOpDone:%i\n", (*inst_list_it)->memOpDone);
        }

        cprintf("\n");

        inst_list_it--;
        ++num;
    }

    cprintf("Waiting list size: %i\n", waitingList.size());

    inst_list_it = --(waitingList.end());

    while (inst_list_it != waitingList.end())
    {
        cprintf("Instruction:%i\n",
                num);
        if (!(*inst_list_it)->isSquashed()) {
            if (!(*inst_list_it)->isIssued()) {
                ++valid_num;
                cprintf("Count:%i\n", valid_num);
            } else if ((*inst_list_it)->isMemRef() &&
                       !(*inst_list_it)->memOpDone) {
                // Loads that have not been marked as executed still count
                // towards the total instructions.
                ++valid_num;
                cprintf("Count:%i\n", valid_num);
            }
        }

        cprintf("PC:%#x\n[sn:%lli]\n[tid:%i]\n"
                "Issued:%i\nSquashed:%i\n",
                (*inst_list_it)->readPC(),
                (*inst_list_it)->seqNum,
                (*inst_list_it)->threadNumber,
                (*inst_list_it)->isIssued(),
                (*inst_list_it)->isSquashed());

        if ((*inst_list_it)->isMemRef()) {
            cprintf("MemOpDone:%i\n", (*inst_list_it)->memOpDone);
        }

        cprintf("\n");

        inst_list_it--;
        ++num;
    }

    cprintf("waitingMemOps list size: %i\n", waitingMemOps.size());

    MemIt waiting_it = waitingMemOps.begin();

    while (waiting_it != waitingMemOps.end())
    {
        cprintf("[sn:%lli] ", (*waiting_it));
        waiting_it++;
        ++num;
    }
    cprintf("\n");
}