sanchayanmaity/gem5
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
gem5/src/cpu/minor/execute.cc
Mitch Hayenga 752f1c1fe9 cpu: Fix Minor SMT WFI/drain interaction issues 
The behavior of WFI is to cause minor to cease evaluating
pipeline logic until an interrupt is observed, however
a user may wish to drain the system while a core is sleeping
due to a WFI.  This patch makes WFI drain.  If an actual
drain occurs during a WFI, the CPU is already drained and will
immediately be ready for swapping, checkpointing, etc.  This
should not negatively impact performance as WFI instructions
are 'stream-changing' (treated like unpredicted branches), so
all remaining instructions are wrong-path and will be squashed
rapidly.

Change-Id: I63833d5acb53d8dde78f9f0c9611de0ece385e45
2016-07-21 17:19:16 +01:00

1865 lines
68 KiB
C++
Raw Blame History

/*
* Copyright (c) 2013-2014 ARM Limited
* All rights reserved
*
* The license below extends only to copyright in the software and shall
* not be construed as granting a license to any other intellectual
* property including but not limited to intellectual property relating
* to a hardware implementation of the functionality of the software
* licensed hereunder. You may use the software subject to the license
* terms below provided that you ensure that this notice is replicated
* unmodified and in its entirety in all distributions of the software,
* modified or unmodified, in source code or in binary form.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are
* met: redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer;
* redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution;
* neither the name of the copyright holders nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* Authors: Andrew Bardsley
*/
#include "arch/locked_mem.hh"
#include "arch/registers.hh"
#include "arch/utility.hh"
#include "cpu/minor/cpu.hh"
#include "cpu/minor/exec_context.hh"
#include "cpu/minor/execute.hh"
#include "cpu/minor/fetch1.hh"
#include "cpu/minor/lsq.hh"
#include "cpu/op_class.hh"
#include "debug/Activity.hh"
#include "debug/Branch.hh"
#include "debug/Drain.hh"
#include "debug/MinorExecute.hh"
#include "debug/MinorInterrupt.hh"
#include "debug/MinorMem.hh"
#include "debug/MinorTrace.hh"
#include "debug/PCEvent.hh"
namespace Minor
{
Execute::Execute(const std::string &name_,
MinorCPU &cpu_,
MinorCPUParams &params,
Latch<ForwardInstData>::Output inp_,
Latch<BranchData>::Input out_) :
Named(name_),
inp(inp_),
out(out_),
cpu(cpu_),
issueLimit(params.executeIssueLimit),
memoryIssueLimit(params.executeMemoryIssueLimit),
commitLimit(params.executeCommitLimit),
memoryCommitLimit(params.executeMemoryCommitLimit),
processMoreThanOneInput(params.executeCycleInput),
fuDescriptions(*params.executeFuncUnits),
numFuncUnits(fuDescriptions.funcUnits.size()),
setTraceTimeOnCommit(params.executeSetTraceTimeOnCommit),
setTraceTimeOnIssue(params.executeSetTraceTimeOnIssue),
allowEarlyMemIssue(params.executeAllowEarlyMemoryIssue),
noCostFUIndex(fuDescriptions.funcUnits.size() + 1),
lsq(name_ + ".lsq", name_ + ".dcache_port",
cpu_, *this,
params.executeMaxAccessesInMemory,
params.executeMemoryWidth,
params.executeLSQRequestsQueueSize,
params.executeLSQTransfersQueueSize,
params.executeLSQStoreBufferSize,
params.executeLSQMaxStoreBufferStoresPerCycle),
executeInfo(params.numThreads, ExecuteThreadInfo(params.executeCommitLimit)),
interruptPriority(0),
issuePriority(0),
commitPriority(0)
{
if (commitLimit < 1) {
fatal("%s: executeCommitLimit must be >= 1 (%d)\n", name_,
commitLimit);
}
if (issueLimit < 1) {
fatal("%s: executeCommitLimit must be >= 1 (%d)\n", name_,
issueLimit);
}
if (memoryIssueLimit < 1) {
fatal("%s: executeMemoryIssueLimit must be >= 1 (%d)\n", name_,
memoryIssueLimit);
}
if (memoryCommitLimit > commitLimit) {
fatal("%s: executeMemoryCommitLimit (%d) must be <="
" executeCommitLimit (%d)\n",
name_, memoryCommitLimit, commitLimit);
}
if (params.executeInputBufferSize < 1) {
fatal("%s: executeInputBufferSize must be >= 1 (%d)\n", name_,
params.executeInputBufferSize);
}
if (params.executeInputBufferSize < 1) {
fatal("%s: executeInputBufferSize must be >= 1 (%d)\n", name_,
params.executeInputBufferSize);
}
/* This should be large enough to count all the in-FU instructions
* which need to be accounted for in the inFlightInsts
* queue */
unsigned int total_slots = 0;
/* Make FUPipelines for each MinorFU */
for (unsigned int i = 0; i < numFuncUnits; i++) {
std::ostringstream fu_name;
MinorFU *fu_description = fuDescriptions.funcUnits[i];
/* Note the total number of instruction slots (for sizing
* the inFlightInst queue) and the maximum latency of any FU
* (for sizing the activity recorder) */
total_slots += fu_description->opLat;
fu_name << name_ << ".fu." << i;
FUPipeline *fu = new FUPipeline(fu_name.str(), *fu_description, cpu);
funcUnits.push_back(fu);
}
/** Check that there is a functional unit for all operation classes */
for (int op_class = No_OpClass + 1; op_class < Num_OpClasses; op_class++) {
bool found_fu = false;
unsigned int fu_index = 0;
while (fu_index < numFuncUnits && !found_fu)
{
if (funcUnits[fu_index]->provides(
static_cast<OpClass>(op_class)))
{
found_fu = true;
}
fu_index++;
}
if (!found_fu) {
warn("No functional unit for OpClass %s\n",
Enums::OpClassStrings[op_class]);
}
}
/* Per-thread structures */
for (ThreadID tid = 0; tid < params.numThreads; tid++) {
std::string tid_str = std::to_string(tid);
/* Input Buffers */
inputBuffer.push_back(
InputBuffer<ForwardInstData>(
name_ + ".inputBuffer" + tid_str, "insts",
params.executeInputBufferSize));
/* Scoreboards */
scoreboard.push_back(Scoreboard(name_ + ".scoreboard" + tid_str));
/* In-flight instruction records */
executeInfo[tid].inFlightInsts = new Queue<QueuedInst,
ReportTraitsAdaptor<QueuedInst> >(
name_ + ".inFlightInsts" + tid_str, "insts", total_slots);
executeInfo[tid].inFUMemInsts = new Queue<QueuedInst,
ReportTraitsAdaptor<QueuedInst> >(
name_ + ".inFUMemInsts" + tid_str, "insts", total_slots);
}
}
const ForwardInstData *
Execute::getInput(ThreadID tid)
{
/* Get a line from the inputBuffer to work with */
if (!inputBuffer[tid].empty()) {
const ForwardInstData &head = inputBuffer[tid].front();
return (head.isBubble() ? NULL : &(inputBuffer[tid].front()));
} else {
return NULL;
}
}
void
Execute::popInput(ThreadID tid)
{
if (!inputBuffer[tid].empty())
inputBuffer[tid].pop();
executeInfo[tid].inputIndex = 0;
}
void
Execute::tryToBranch(MinorDynInstPtr inst, Fault fault, BranchData &branch)
{
ThreadContext *thread = cpu.getContext(inst->id.threadId);
const TheISA::PCState &pc_before = inst->pc;
TheISA::PCState target = thread->pcState();
/* Force a branch for SerializeAfter instructions at the end of micro-op
* sequence when we're not suspended */
bool force_branch = thread->status() != ThreadContext::Suspended &&
!inst->isFault() &&
inst->isLastOpInInst() &&
(inst->staticInst->isSerializeAfter() ||
inst->staticInst->isIprAccess());
DPRINTF(Branch, "tryToBranch before: %s after: %s%s\n",
pc_before, target, (force_branch ? " (forcing)" : ""));
/* Will we change the PC to something other than the next instruction? */
bool must_branch = pc_before != target ||
fault != NoFault ||
force_branch;
/* The reason for the branch data we're about to generate, set below */
BranchData::Reason reason = BranchData::NoBranch;
if (fault == NoFault)
{
TheISA::advancePC(target, inst->staticInst);
thread->pcState(target);
DPRINTF(Branch, "Advancing current PC from: %s to: %s\n",
pc_before, target);
}
if (inst->predictedTaken && !force_branch) {
/* Predicted to branch */
if (!must_branch) {
/* No branch was taken, change stream to get us back to the
* intended PC value */
DPRINTF(Branch, "Predicted a branch from 0x%x to 0x%x but"
" none happened inst: %s\n",
inst->pc.instAddr(), inst->predictedTarget.instAddr(), *inst);
reason = BranchData::BadlyPredictedBranch;
} else if (inst->predictedTarget == target) {
/* Branch prediction got the right target, kill the branch and
* carry on.
* Note that this information to the branch predictor might get
* overwritten by a "real" branch during this cycle */
DPRINTF(Branch, "Predicted a branch from 0x%x to 0x%x correctly"
" inst: %s\n",
inst->pc.instAddr(), inst->predictedTarget.instAddr(), *inst);
reason = BranchData::CorrectlyPredictedBranch;
} else {
/* Branch prediction got the wrong target */
DPRINTF(Branch, "Predicted a branch from 0x%x to 0x%x"
" but got the wrong target (actual: 0x%x) inst: %s\n",
inst->pc.instAddr(), inst->predictedTarget.instAddr(),
target.instAddr(), *inst);
reason = BranchData::BadlyPredictedBranchTarget;
}
} else if (must_branch) {
/* Unpredicted branch */
DPRINTF(Branch, "Unpredicted branch from 0x%x to 0x%x inst: %s\n",
inst->pc.instAddr(), target.instAddr(), *inst);
reason = BranchData::UnpredictedBranch;
} else {
/* No branch at all */
reason = BranchData::NoBranch;
}
updateBranchData(inst->id.threadId, reason, inst, target, branch);
}
void
Execute::updateBranchData(
ThreadID tid,
BranchData::Reason reason,
MinorDynInstPtr inst, const TheISA::PCState &target,
BranchData &branch)
{
if (reason != BranchData::NoBranch) {
/* Bump up the stream sequence number on a real branch*/
if (BranchData::isStreamChange(reason))
executeInfo[tid].streamSeqNum++;
/* Branches (even mis-predictions) don't change the predictionSeqNum,
* just the streamSeqNum */
branch = BranchData(reason, tid,
executeInfo[tid].streamSeqNum,
/* Maintaining predictionSeqNum if there's no inst is just a
* courtesy and looks better on minorview */
(inst->isBubble() ? executeInfo[tid].lastPredictionSeqNum
: inst->id.predictionSeqNum),
target, inst);
DPRINTF(Branch, "Branch data signalled: %s\n", branch);
}
}
void
Execute::handleMemResponse(MinorDynInstPtr inst,
LSQ::LSQRequestPtr response, BranchData &branch, Fault &fault)
{
ThreadID thread_id = inst->id.threadId;
ThreadContext *thread = cpu.getContext(thread_id);
ExecContext context(cpu, *cpu.threads[thread_id], *this, inst);
PacketPtr packet = response->packet;
bool is_load = inst->staticInst->isLoad();
bool is_store = inst->staticInst->isStore();
bool is_prefetch = inst->staticInst->isDataPrefetch();
/* If true, the trace's predicate value will be taken from the exec
* context predicate, otherwise, it will be set to false */
bool use_context_predicate = true;
if (response->fault != NoFault) {
/* Invoke memory faults. */
DPRINTF(MinorMem, "Completing fault from DTLB access: %s\n",
response->fault->name());
if (inst->staticInst->isPrefetch()) {
DPRINTF(MinorMem, "Not taking fault on prefetch: %s\n",
response->fault->name());
/* Don't assign to fault */
} else {
/* Take the fault raised during the TLB/memory access */
fault = response->fault;
fault->invoke(thread, inst->staticInst);
}
} else if (!packet) {
DPRINTF(MinorMem, "Completing failed request inst: %s\n",
*inst);
use_context_predicate = false;
} else if (packet->isError()) {
DPRINTF(MinorMem, "Trying to commit error response: %s\n",
*inst);
fatal("Received error response packet for inst: %s\n", *inst);
} else if (is_store || is_load || is_prefetch) {
assert(packet);
DPRINTF(MinorMem, "Memory response inst: %s addr: 0x%x size: %d\n",
*inst, packet->getAddr(), packet->getSize());
if (is_load && packet->getSize() > 0) {
DPRINTF(MinorMem, "Memory data[0]: 0x%x\n",
static_cast<unsigned int>(packet->getConstPtr<uint8_t>()[0]));
}
/* Complete the memory access instruction */
fault = inst->staticInst->completeAcc(packet, &context,
inst->traceData);
if (fault != NoFault) {
/* Invoke fault created by instruction completion */
DPRINTF(MinorMem, "Fault in memory completeAcc: %s\n",
fault->name());
fault->invoke(thread, inst->staticInst);
} else {
/* Stores need to be pushed into the store buffer to finish
* them off */
if (response->needsToBeSentToStoreBuffer())
lsq.sendStoreToStoreBuffer(response);
}
} else {
fatal("There should only ever be reads, "
"writes or faults at this point\n");
}
lsq.popResponse(response);
if (inst->traceData) {
inst->traceData->setPredicate((use_context_predicate ?
context.readPredicate() : false));
}
doInstCommitAccounting(inst);
/* Generate output to account for branches */
tryToBranch(inst, fault, branch);
}
bool
Execute::isInterrupted(ThreadID thread_id) const
{
return cpu.checkInterrupts(cpu.getContext(thread_id));
}
bool
Execute::takeInterrupt(ThreadID thread_id, BranchData &branch)
{
DPRINTF(MinorInterrupt, "Considering interrupt status from PC: %s\n",
cpu.getContext(thread_id)->pcState());
Fault interrupt = cpu.getInterruptController(thread_id)->getInterrupt
(cpu.getContext(thread_id));
if (interrupt != NoFault) {
/* The interrupt *must* set pcState */
cpu.getInterruptController(thread_id)->updateIntrInfo
(cpu.getContext(thread_id));
interrupt->invoke(cpu.getContext(thread_id));
assert(!lsq.accessesInFlight());
DPRINTF(MinorInterrupt, "Invoking interrupt: %s to PC: %s\n",
interrupt->name(), cpu.getContext(thread_id)->pcState());
/* Assume that an interrupt *must* cause a branch. Assert this? */
updateBranchData(thread_id, BranchData::Interrupt,
MinorDynInst::bubble(), cpu.getContext(thread_id)->pcState(),
branch);
}
return interrupt != NoFault;
}
bool
Execute::executeMemRefInst(MinorDynInstPtr inst, BranchData &branch,
bool &passed_predicate, Fault &fault)
{
bool issued = false;
/* Set to true if the mem op. is issued and sent to the mem system */
passed_predicate = false;
if (!lsq.canRequest()) {
/* Not acting on instruction yet as the memory
* queues are full */
issued = false;
} else {
ThreadContext *thread = cpu.getContext(inst->id.threadId);
TheISA::PCState old_pc = thread->pcState();
ExecContext context(cpu, *cpu.threads[inst->id.threadId],
*this, inst);
DPRINTF(MinorExecute, "Initiating memRef inst: %s\n", *inst);
Fault init_fault = inst->staticInst->initiateAcc(&context,
inst->traceData);
if (init_fault != NoFault) {
DPRINTF(MinorExecute, "Fault on memory inst: %s"
" initiateAcc: %s\n", *inst, init_fault->name());
fault = init_fault;
} else {
/* Only set this if the instruction passed its
* predicate */
passed_predicate = context.readPredicate();
/* Set predicate in tracing */
if (inst->traceData)
inst->traceData->setPredicate(passed_predicate);
/* If the instruction didn't pass its predicate (and so will not
* progress from here) Try to branch to correct and branch
* mis-prediction. */
if (!passed_predicate) {
/* Leave it up to commit to handle the fault */
lsq.pushFailedRequest(inst);
}
}
/* Restore thread PC */
thread->pcState(old_pc);
issued = true;
}
return issued;
}
/** Increment a cyclic buffer index for indices [0, cycle_size-1] */
inline unsigned int
cyclicIndexInc(unsigned int index, unsigned int cycle_size)
{
unsigned int ret = index + 1;
if (ret == cycle_size)
ret = 0;
return ret;
}
/** Decrement a cyclic buffer index for indices [0, cycle_size-1] */
inline unsigned int
cyclicIndexDec(unsigned int index, unsigned int cycle_size)
{
int ret = index - 1;
if (ret < 0)
ret = cycle_size - 1;
return ret;
}
unsigned int
Execute::issue(ThreadID thread_id)
{
const ForwardInstData *insts_in = getInput(thread_id);
ExecuteThreadInfo &thread = executeInfo[thread_id];
/* Early termination if we have no instructions */
if (!insts_in)
return 0;
/* Start from the first FU */
unsigned int fu_index = 0;
/* Remains true while instructions are still being issued. If any
* instruction fails to issue, this is set to false and we exit issue.
* This strictly enforces in-order issue. For other issue behaviours,
* a more complicated test in the outer while loop below is needed. */
bool issued = true;
/* Number of insts issues this cycle to check for issueLimit */
unsigned num_insts_issued = 0;
/* Number of memory ops issues this cycle to check for memoryIssueLimit */
unsigned num_mem_insts_issued = 0;
/* Number of instructions discarded this cycle in order to enforce a
* discardLimit. @todo, add that parameter? */
unsigned num_insts_discarded = 0;
do {
MinorDynInstPtr inst = insts_in->insts[thread.inputIndex];
Fault fault = inst->fault;
bool discarded = false;
bool issued_mem_ref = false;
if (inst->isBubble()) {
/* Skip */
issued = true;
} else if (cpu.getContext(thread_id)->status() ==
ThreadContext::Suspended)
{
DPRINTF(MinorExecute, "Discarding inst: %s from suspended"
" thread\n", *inst);
issued = true;
discarded = true;
} else if (inst->id.streamSeqNum != thread.streamSeqNum) {
DPRINTF(MinorExecute, "Discarding inst: %s as its stream"
" state was unexpected, expected: %d\n",
*inst, thread.streamSeqNum);
issued = true;
discarded = true;
} else {
/* Try and issue an instruction into an FU, assume we didn't and
* fix that in the loop */
issued = false;
/* Try FU from 0 each instruction */
fu_index = 0;
/* Try and issue a single instruction stepping through the
* available FUs */
do {
FUPipeline *fu = funcUnits[fu_index];
DPRINTF(MinorExecute, "Trying to issue inst: %s to FU: %d\n",
*inst, fu_index);
/* Does the examined fu have the OpClass-related capability
* needed to execute this instruction? Faults can always
* issue to any FU but probably should just 'live' in the
* inFlightInsts queue rather than having an FU. */
bool fu_is_capable = (!inst->isFault() ?
fu->provides(inst->staticInst->opClass()) : true);
if (inst->isNoCostInst()) {
/* Issue free insts. to a fake numbered FU */
fu_index = noCostFUIndex;
/* And start the countdown on activity to allow
* this instruction to get to the end of its FU */
cpu.activityRecorder->activity();
/* Mark the destinations for this instruction as
* busy */
scoreboard[thread_id].markupInstDests(inst, cpu.curCycle() +
Cycles(0), cpu.getContext(thread_id), false);
DPRINTF(MinorExecute, "Issuing %s to %d\n", inst->id, noCostFUIndex);
inst->fuIndex = noCostFUIndex;
inst->extraCommitDelay = Cycles(0);
inst->extraCommitDelayExpr = NULL;
/* Push the instruction onto the inFlight queue so
* it can be committed in order */
QueuedInst fu_inst(inst);
thread.inFlightInsts->push(fu_inst);
issued = true;
} else if (!fu_is_capable || fu->alreadyPushed()) {
/* Skip */
if (!fu_is_capable) {
DPRINTF(MinorExecute, "Can't issue as FU: %d isn't"
" capable\n", fu_index);
} else {
DPRINTF(MinorExecute, "Can't issue as FU: %d is"
" already busy\n", fu_index);
}
} else if (fu->stalled) {
DPRINTF(MinorExecute, "Can't issue inst: %s into FU: %d,"
" it's stalled\n",
*inst, fu_index);
} else if (!fu->canInsert()) {
DPRINTF(MinorExecute, "Can't issue inst: %s to busy FU"
" for another: %d cycles\n",
*inst, fu->cyclesBeforeInsert());
} else {
MinorFUTiming *timing = (!inst->isFault() ?
fu->findTiming(inst->staticInst) : NULL);
const std::vector<Cycles> *src_latencies =
(timing ? &(timing->srcRegsRelativeLats)
: NULL);
const std::vector<bool> *cant_forward_from_fu_indices =
&(fu->cantForwardFromFUIndices);
if (timing && timing->suppress) {
DPRINTF(MinorExecute, "Can't issue inst: %s as extra"
" decoding is suppressing it\n",
*inst);
} else if (!scoreboard[thread_id].canInstIssue(inst,
src_latencies, cant_forward_from_fu_indices,
cpu.curCycle(), cpu.getContext(thread_id)))
{
DPRINTF(MinorExecute, "Can't issue inst: %s yet\n",
*inst);
} else {
/* Can insert the instruction into this FU */
DPRINTF(MinorExecute, "Issuing inst: %s"
" into FU %d\n", *inst,
fu_index);
Cycles extra_dest_retire_lat = Cycles(0);
TimingExpr *extra_dest_retire_lat_expr = NULL;
Cycles extra_assumed_lat = Cycles(0);
/* Add the extraCommitDelay and extraAssumeLat to
* the FU pipeline timings */
if (timing) {
extra_dest_retire_lat =
timing->extraCommitLat;
extra_dest_retire_lat_expr =
timing->extraCommitLatExpr;
extra_assumed_lat =
timing->extraAssumedLat;
}
issued_mem_ref = inst->isMemRef();
QueuedInst fu_inst(inst);
/* Decorate the inst with FU details */
inst->fuIndex = fu_index;
inst->extraCommitDelay = extra_dest_retire_lat;
inst->extraCommitDelayExpr =
extra_dest_retire_lat_expr;
if (issued_mem_ref) {
/* Remember which instruction this memory op
* depends on so that initiateAcc can be called
* early */
if (allowEarlyMemIssue) {
inst->instToWaitFor =
scoreboard[thread_id].execSeqNumToWaitFor(inst,
cpu.getContext(thread_id));
if (lsq.getLastMemBarrier(thread_id) >
inst->instToWaitFor)
{
DPRINTF(MinorExecute, "A barrier will"
" cause a delay in mem ref issue of"
" inst: %s until after inst"
" %d(exec)\n", *inst,
lsq.getLastMemBarrier(thread_id));
inst->instToWaitFor =
lsq.getLastMemBarrier(thread_id);
} else {
DPRINTF(MinorExecute, "Memory ref inst:"
" %s must wait for inst %d(exec)"
" before issuing\n",
*inst, inst->instToWaitFor);
}
inst->canEarlyIssue = true;
}
/* Also queue this instruction in the memory ref
* queue to ensure in-order issue to the LSQ */
DPRINTF(MinorExecute, "Pushing mem inst: %s\n",
*inst);
thread.inFUMemInsts->push(fu_inst);
}
/* Issue to FU */
fu->push(fu_inst);
/* And start the countdown on activity to allow
* this instruction to get to the end of its FU */
cpu.activityRecorder->activity();
/* Mark the destinations for this instruction as
* busy */
scoreboard[thread_id].markupInstDests(inst, cpu.curCycle() +
fu->description.opLat +
extra_dest_retire_lat +
extra_assumed_lat,
cpu.getContext(thread_id),
issued_mem_ref && extra_assumed_lat == Cycles(0));
/* Push the instruction onto the inFlight queue so
* it can be committed in order */
thread.inFlightInsts->push(fu_inst);
issued = true;
}
}
fu_index++;
} while (fu_index != numFuncUnits && !issued);
if (!issued)
DPRINTF(MinorExecute, "Didn't issue inst: %s\n", *inst);
}
if (issued) {
/* Generate MinorTrace's MinorInst lines. Do this at commit
* to allow better instruction annotation? */
if (DTRACE(MinorTrace) && !inst->isBubble())
inst->minorTraceInst(*this);
/* Mark up barriers in the LSQ */
if (!discarded && inst->isInst() &&
inst->staticInst->isMemBarrier())
{
DPRINTF(MinorMem, "Issuing memory barrier inst: %s\n", *inst);
lsq.issuedMemBarrierInst(inst);
}
if (inst->traceData && setTraceTimeOnIssue) {
inst->traceData->setWhen(curTick());
}
if (issued_mem_ref)
num_mem_insts_issued++;
if (discarded) {
num_insts_discarded++;
} else if (!inst->isBubble()) {
num_insts_issued++;
if (num_insts_issued == issueLimit)
DPRINTF(MinorExecute, "Reached inst issue limit\n");
}
thread.inputIndex++;
DPRINTF(MinorExecute, "Stepping to next inst inputIndex: %d\n",
thread.inputIndex);
}
/* Got to the end of a line */
if (thread.inputIndex == insts_in->width()) {
popInput(thread_id);
/* Set insts_in to null to force us to leave the surrounding
* loop */
insts_in = NULL;
if (processMoreThanOneInput) {
DPRINTF(MinorExecute, "Wrapping\n");
insts_in = getInput(thread_id);
}
}
} while (insts_in && thread.inputIndex < insts_in->width() &&
/* We still have instructions */
fu_index != numFuncUnits && /* Not visited all FUs */
issued && /* We've not yet failed to issue an instruction */
num_insts_issued != issueLimit && /* Still allowed to issue */
num_mem_insts_issued != memoryIssueLimit);
return num_insts_issued;
}
bool
Execute::tryPCEvents(ThreadID thread_id)
{
ThreadContext *thread = cpu.getContext(thread_id);
unsigned int num_pc_event_checks = 0;
/* Handle PC events on instructions */
Addr oldPC;
do {
oldPC = thread->instAddr();
cpu.system->pcEventQueue.service(thread);
num_pc_event_checks++;
} while (oldPC != thread->instAddr());
if (num_pc_event_checks > 1) {
DPRINTF(PCEvent, "Acting on PC Event to PC: %s\n",
thread->pcState());
}
return num_pc_event_checks > 1;
}
void
Execute::doInstCommitAccounting(MinorDynInstPtr inst)
{
assert(!inst->isFault());
MinorThread *thread = cpu.threads[inst->id.threadId];
/* Increment the many and various inst and op counts in the
* thread and system */
if (!inst->staticInst->isMicroop() || inst->staticInst->isLastMicroop())
{
thread->numInst++;
thread->numInsts++;
cpu.stats.numInsts++;
cpu.system->totalNumInsts++;
/* Act on events related to instruction counts */
cpu.comInstEventQueue[inst->id.threadId]->serviceEvents(thread->numInst);
cpu.system->instEventQueue.serviceEvents(cpu.system->totalNumInsts);
}
thread->numOp++;
thread->numOps++;
cpu.stats.numOps++;
cpu.stats.committedInstType[inst->id.threadId]
[inst->staticInst->opClass()]++;
/* Set the CP SeqNum to the numOps commit number */
if (inst->traceData)
inst->traceData->setCPSeq(thread->numOp);
cpu.probeInstCommit(inst->staticInst);
}
bool
Execute::commitInst(MinorDynInstPtr inst, bool early_memory_issue,
BranchData &branch, Fault &fault, bool &committed,
bool &completed_mem_issue)
{
ThreadID thread_id = inst->id.threadId;
ThreadContext *thread = cpu.getContext(thread_id);
bool completed_inst = true;
fault = NoFault;
/* Is the thread for this instruction suspended? In that case, just
* stall as long as there are no pending interrupts */
if (thread->status() == ThreadContext::Suspended &&
!isInterrupted(thread_id))
{
panic("We should never hit the case where we try to commit from a "
"suspended thread as the streamSeqNum should not match");
} else if (inst->isFault()) {
ExecContext context(cpu, *cpu.threads[thread_id], *this, inst);
DPRINTF(MinorExecute, "Fault inst reached Execute: %s\n",
inst->fault->name());
fault = inst->fault;
inst->fault->invoke(thread, NULL);
tryToBranch(inst, fault, branch);
} else if (inst->staticInst->isMemRef()) {
/* Memory accesses are executed in two parts:
* executeMemRefInst -- calculates the EA and issues the access
* to memory. This is done here.
* handleMemResponse -- handles the response packet, done by
* Execute::commit
*
* While the memory access is in its FU, the EA is being
* calculated. At the end of the FU, when it is ready to
* 'commit' (in this function), the access is presented to the
* memory queues. When a response comes back from memory,
* Execute::commit will commit it.
*/
bool predicate_passed = false;
bool completed_mem_inst = executeMemRefInst(inst, branch,
predicate_passed, fault);
if (completed_mem_inst && fault != NoFault) {
if (early_memory_issue) {
DPRINTF(MinorExecute, "Fault in early executing inst: %s\n",
fault->name());
/* Don't execute the fault, just stall the instruction
* until it gets to the head of inFlightInsts */
inst->canEarlyIssue = false;
/* Not completed as we'll come here again to pick up
* the fault when we get to the end of the FU */
completed_inst = false;
} else {
DPRINTF(MinorExecute, "Fault in execute: %s\n",
fault->name());
fault->invoke(thread, NULL);
tryToBranch(inst, fault, branch);
completed_inst = true;
}
} else {
completed_inst = completed_mem_inst;
}
completed_mem_issue = completed_inst;
} else if (inst->isInst() && inst->staticInst->isMemBarrier() &&
!lsq.canPushIntoStoreBuffer())
{
DPRINTF(MinorExecute, "Can't commit data barrier inst: %s yet as"
" there isn't space in the store buffer\n", *inst);
completed_inst = false;
} else if (inst->isInst() && inst->staticInst->isQuiesce()
&& !branch.isBubble()){
/* This instruction can suspend, need to be able to communicate
* backwards, so no other branches may evaluate this cycle*/
completed_inst = false;
} else {
ExecContext context(cpu, *cpu.threads[thread_id], *this, inst);
DPRINTF(MinorExecute, "Committing inst: %s\n", *inst);
fault = inst->staticInst->execute(&context,
inst->traceData);
/* Set the predicate for tracing and dump */
if (inst->traceData)
inst->traceData->setPredicate(context.readPredicate());
committed = true;
if (fault != NoFault) {
DPRINTF(MinorExecute, "Fault in execute of inst: %s fault: %s\n",
*inst, fault->name());
fault->invoke(thread, inst->staticInst);
}
doInstCommitAccounting(inst);
tryToBranch(inst, fault, branch);
}
if (completed_inst) {
/* Keep a copy of this instruction's predictionSeqNum just in case
* we need to issue a branch without an instruction (such as an
* interrupt) */
executeInfo[thread_id].lastPredictionSeqNum = inst->id.predictionSeqNum;
/* Check to see if this instruction suspended the current thread. */
if (!inst->isFault() &&
thread->status() == ThreadContext::Suspended &&
branch.isBubble() && /* It didn't branch too */
!isInterrupted(thread_id)) /* Don't suspend if we have
interrupts */
{
TheISA::PCState resume_pc = cpu.getContext(thread_id)->pcState();
assert(resume_pc.microPC() == 0);
DPRINTF(MinorInterrupt, "Suspending thread: %d from Execute"
" inst: %s\n", thread_id, *inst);
cpu.stats.numFetchSuspends++;
updateBranchData(thread_id, BranchData::SuspendThread, inst,
resume_pc, branch);
}
}
return completed_inst;
}
void
Execute::commit(ThreadID thread_id, bool only_commit_microops, bool discard,
BranchData &branch)
{
Fault fault = NoFault;
Cycles now = cpu.curCycle();
ExecuteThreadInfo &ex_info = executeInfo[thread_id];
/**
* Try and execute as many instructions from the end of FU pipelines as
* possible. This *doesn't* include actually advancing the pipelines.
*
* We do this by looping on the front of the inFlightInsts queue for as
* long as we can find the desired instruction at the end of the
* functional unit it was issued to without seeing a branch or a fault.
* In this function, these terms are used:
* complete -- The instruction has finished its passage through
* its functional unit and its fate has been decided
* (committed, discarded, issued to the memory system)
* commit -- The instruction is complete(d), not discarded and has
* its effects applied to the CPU state
* discard(ed) -- The instruction is complete but not committed
* as its streamSeqNum disagrees with the current
* Execute::streamSeqNum
*
* Commits are also possible from two other places:
*
* 1) Responses returning from the LSQ
* 2) Mem ops issued to the LSQ ('committed' from the FUs) earlier
* than their position in the inFlightInsts queue, but after all
* their dependencies are resolved.
*/
/* Has an instruction been completed? Once this becomes false, we stop
* trying to complete instructions. */
bool completed_inst = true;
/* Number of insts committed this cycle to check against commitLimit */
unsigned int num_insts_committed = 0;
/* Number of memory access instructions committed to check against
* memCommitLimit */
unsigned int num_mem_refs_committed = 0;
if (only_commit_microops && !ex_info.inFlightInsts->empty()) {
DPRINTF(MinorInterrupt, "Only commit microops %s %d\n",
*(ex_info.inFlightInsts->front().inst),
ex_info.lastCommitWasEndOfMacroop);
}
while (!ex_info.inFlightInsts->empty() && /* Some more instructions to process */
!branch.isStreamChange() && /* No real branch */
fault == NoFault && /* No faults */
completed_inst && /* Still finding instructions to execute */
num_insts_committed != commitLimit /* Not reached commit limit */
)
{
if (only_commit_microops) {
DPRINTF(MinorInterrupt, "Committing tail of insts before"
" interrupt: %s\n",
*(ex_info.inFlightInsts->front().inst));
}
QueuedInst *head_inflight_inst = &(ex_info.inFlightInsts->front());
InstSeqNum head_exec_seq_num =
head_inflight_inst->inst->id.execSeqNum;
/* The instruction we actually process if completed_inst
* remains true to the end of the loop body.
* Start by considering the the head of the in flight insts queue */
MinorDynInstPtr inst = head_inflight_inst->inst;
bool committed_inst = false;
bool discard_inst = false;
bool completed_mem_ref = false;
bool issued_mem_ref = false;
bool early_memory_issue = false;
/* Must set this again to go around the loop */
completed_inst = false;
/* If we're just completing a macroop before an interrupt or drain,
* can we stil commit another microop (rather than a memory response)
* without crosing into the next full instruction? */
bool can_commit_insts = !ex_info.inFlightInsts->empty() &&
!(only_commit_microops && ex_info.lastCommitWasEndOfMacroop);
/* Can we find a mem response for this inst */
LSQ::LSQRequestPtr mem_response =
(inst->inLSQ ? lsq.findResponse(inst) : NULL);
DPRINTF(MinorExecute, "Trying to commit canCommitInsts: %d\n",
can_commit_insts);
/* Test for PC events after every instruction */
if (isInbetweenInsts(thread_id) && tryPCEvents(thread_id)) {
ThreadContext *thread = cpu.getContext(thread_id);
/* Branch as there was a change in PC */
updateBranchData(thread_id, BranchData::UnpredictedBranch,
MinorDynInst::bubble(), thread->pcState(), branch);
} else if (mem_response &&
num_mem_refs_committed < memoryCommitLimit)
{
/* Try to commit from the memory responses next */
discard_inst = inst->id.streamSeqNum !=
ex_info.streamSeqNum || discard;
DPRINTF(MinorExecute, "Trying to commit mem response: %s\n",
*inst);
/* Complete or discard the response */
if (discard_inst) {
DPRINTF(MinorExecute, "Discarding mem inst: %s as its"
" stream state was unexpected, expected: %d\n",
*inst, ex_info.streamSeqNum);
lsq.popResponse(mem_response);
} else {
handleMemResponse(inst, mem_response, branch, fault);
committed_inst = true;
}
completed_mem_ref = true;
completed_inst = true;
} else if (can_commit_insts) {
/* If true, this instruction will, subject to timing tweaks,
* be considered for completion. try_to_commit flattens
* the `if' tree a bit and allows other tests for inst
* commit to be inserted here. */
bool try_to_commit = false;
/* Try and issue memory ops early if they:
* - Can push a request into the LSQ
* - Have reached the end of their FUs
* - Have had all their dependencies satisfied
* - Are from the right stream
*
* For any other case, leave it to the normal instruction
* issue below to handle them.
*/
if (!ex_info.inFUMemInsts->empty() && lsq.canRequest()) {
DPRINTF(MinorExecute, "Trying to commit from mem FUs\n");
const MinorDynInstPtr head_mem_ref_inst =
ex_info.inFUMemInsts->front().inst;
FUPipeline *fu = funcUnits[head_mem_ref_inst->fuIndex];
const MinorDynInstPtr &fu_inst = fu->front().inst;
/* Use this, possibly out of order, inst as the one
* to 'commit'/send to the LSQ */
if (!fu_inst->isBubble() &&
!fu_inst->inLSQ &&
fu_inst->canEarlyIssue &&
ex_info.streamSeqNum == fu_inst->id.streamSeqNum &&
head_exec_seq_num > fu_inst->instToWaitFor)
{
DPRINTF(MinorExecute, "Issuing mem ref early"
" inst: %s instToWaitFor: %d\n",
*(fu_inst), fu_inst->instToWaitFor);
inst = fu_inst;
try_to_commit = true;
early_memory_issue = true;
completed_inst = true;
}
}
/* Try and commit FU-less insts */
if (!completed_inst && inst->isNoCostInst()) {
DPRINTF(MinorExecute, "Committing no cost inst: %s", *inst);
try_to_commit = true;
completed_inst = true;
}
/* Try to issue from the ends of FUs and the inFlightInsts
* queue */
if (!completed_inst && !inst->inLSQ) {
DPRINTF(MinorExecute, "Trying to commit from FUs\n");
/* Try to commit from a functional unit */
/* Is the head inst of the expected inst's FU actually the
* expected inst? */
QueuedInst &fu_inst =
funcUnits[inst->fuIndex]->front();
InstSeqNum fu_inst_seq_num = fu_inst.inst->id.execSeqNum;
if (fu_inst.inst->isBubble()) {
/* No instruction ready */
completed_inst = false;
} else if (fu_inst_seq_num != head_exec_seq_num) {
/* Past instruction: we must have already executed it
* in the same cycle and so the head inst isn't
* actually at the end of its pipeline
* Future instruction: handled above and only for
* mem refs on their way to the LSQ */
} else if (fu_inst.inst->id == inst->id) {
/* All instructions can be committed if they have the
* right execSeqNum and there are no in-flight
* mem insts before us */
try_to_commit = true;
completed_inst = true;
}
}
if (try_to_commit) {
discard_inst = inst->id.streamSeqNum !=
ex_info.streamSeqNum || discard;
/* Is this instruction discardable as its streamSeqNum
* doesn't match? */
if (!discard_inst) {
/* Try to commit or discard a non-memory instruction.
* Memory ops are actually 'committed' from this FUs
* and 'issued' into the memory system so we need to
* account for them later (commit_was_mem_issue gets
* set) */
if (inst->extraCommitDelayExpr) {
DPRINTF(MinorExecute, "Evaluating expression for"
" extra commit delay inst: %s\n", *inst);
ThreadContext *thread = cpu.getContext(thread_id);
TimingExprEvalContext context(inst->staticInst,
thread, NULL);
uint64_t extra_delay = inst->extraCommitDelayExpr->
eval(context);
DPRINTF(MinorExecute, "Extra commit delay expr"
" result: %d\n", extra_delay);
if (extra_delay < 128) {
inst->extraCommitDelay += Cycles(extra_delay);
} else {
DPRINTF(MinorExecute, "Extra commit delay was"
" very long: %d\n", extra_delay);
}
inst->extraCommitDelayExpr = NULL;
}
/* Move the extraCommitDelay from the instruction
* into the minimumCommitCycle */
if (inst->extraCommitDelay != Cycles(0)) {
inst->minimumCommitCycle = cpu.curCycle() +
inst->extraCommitDelay;
inst->extraCommitDelay = Cycles(0);
}
/* @todo Think about making lastMemBarrier be
* MAX_UINT_64 to avoid using 0 as a marker value */
if (!inst->isFault() && inst->isMemRef() &&
lsq.getLastMemBarrier(thread_id) <
inst->id.execSeqNum &&
lsq.getLastMemBarrier(thread_id) != 0)
{
DPRINTF(MinorExecute, "Not committing inst: %s yet"
" as there are incomplete barriers in flight\n",
*inst);
completed_inst = false;
} else if (inst->minimumCommitCycle > now) {
DPRINTF(MinorExecute, "Not committing inst: %s yet"
" as it wants to be stalled for %d more cycles\n",
*inst, inst->minimumCommitCycle - now);
completed_inst = false;
} else {
completed_inst = commitInst(inst,
early_memory_issue, branch, fault,
committed_inst, issued_mem_ref);
}
} else {
/* Discard instruction */
completed_inst = true;
}
if (completed_inst) {
/* Allow the pipeline to advance. If the FU head
* instruction wasn't the inFlightInsts head
* but had already been committed, it would have
* unstalled the pipeline before here */
if (inst->fuIndex != noCostFUIndex) {
DPRINTF(MinorExecute, "Unstalling %d for inst %s\n", inst->fuIndex, inst->id);
funcUnits[inst->fuIndex]->stalled = false;
}
}
}
} else {
DPRINTF(MinorExecute, "No instructions to commit\n");
completed_inst = false;
}
/* All discardable instructions must also be 'completed' by now */
assert(!(discard_inst && !completed_inst));
/* Instruction committed but was discarded due to streamSeqNum
* mismatch */
if (discard_inst) {
DPRINTF(MinorExecute, "Discarding inst: %s as its stream"
" state was unexpected, expected: %d\n",
*inst, ex_info.streamSeqNum);
if (fault == NoFault)
cpu.stats.numDiscardedOps++;
}
/* Mark the mem inst as being in the LSQ */
if (issued_mem_ref) {
inst->fuIndex = 0;
inst->inLSQ = true;
}
/* Pop issued (to LSQ) and discarded mem refs from the inFUMemInsts
* as they've *definitely* exited the FUs */
if (completed_inst && inst->isMemRef()) {
/* The MemRef could have been discarded from the FU or the memory
* queue, so just check an FU instruction */
if (!ex_info.inFUMemInsts->empty() &&
ex_info.inFUMemInsts->front().inst == inst)
{
ex_info.inFUMemInsts->pop();
}
}
if (completed_inst && !(issued_mem_ref && fault == NoFault)) {
/* Note that this includes discarded insts */
DPRINTF(MinorExecute, "Completed inst: %s\n", *inst);
/* Got to the end of a full instruction? */
ex_info.lastCommitWasEndOfMacroop = inst->isFault() ||
inst->isLastOpInInst();
/* lastPredictionSeqNum is kept as a convenience to prevent its
* value from changing too much on the minorview display */
ex_info.lastPredictionSeqNum = inst->id.predictionSeqNum;
/* Finished with the inst, remove it from the inst queue and
* clear its dependencies */
ex_info.inFlightInsts->pop();
/* Complete barriers in the LSQ/move to store buffer */
if (inst->isInst() && inst->staticInst->isMemBarrier()) {
DPRINTF(MinorMem, "Completing memory barrier"
" inst: %s committed: %d\n", *inst, committed_inst);
lsq.completeMemBarrierInst(inst, committed_inst);
}
scoreboard[thread_id].clearInstDests(inst, inst->isMemRef());
}
/* Handle per-cycle instruction counting */
if (committed_inst) {
bool is_no_cost_inst = inst->isNoCostInst();
/* Don't show no cost instructions as having taken a commit
* slot */
if (DTRACE(MinorTrace) && !is_no_cost_inst)
ex_info.instsBeingCommitted.insts[num_insts_committed] = inst;
if (!is_no_cost_inst)
num_insts_committed++;
if (num_insts_committed == commitLimit)
DPRINTF(MinorExecute, "Reached inst commit limit\n");
/* Re-set the time of the instruction if that's required for
* tracing */
if (inst->traceData) {
if (setTraceTimeOnCommit)
inst->traceData->setWhen(curTick());
inst->traceData->dump();
}
if (completed_mem_ref)
num_mem_refs_committed++;
if (num_mem_refs_committed == memoryCommitLimit)
DPRINTF(MinorExecute, "Reached mem ref commit limit\n");
}
}
}
bool
Execute::isInbetweenInsts(ThreadID thread_id) const
{
return executeInfo[thread_id].lastCommitWasEndOfMacroop &&
!lsq.accessesInFlight();
}
void
Execute::evaluate()
{
if (!inp.outputWire->isBubble())
inputBuffer[inp.outputWire->threadId].setTail(*inp.outputWire);
BranchData &branch = *out.inputWire;
unsigned int num_issued = 0;
/* Do all the cycle-wise activities for dcachePort here to potentially
* free up input spaces in the LSQ's requests queue */
lsq.step();
/* Check interrupts first. Will halt commit if interrupt found */
bool interrupted = false;
ThreadID interrupt_tid = checkInterrupts(branch, interrupted);
if (interrupt_tid != InvalidThreadID) {
/* Signalling an interrupt this cycle, not issuing/committing from
* any other threads */
} else if (!branch.isBubble()) {
/* It's important that this is here to carry Fetch1 wakeups to Fetch1
* without overwriting them */
DPRINTF(MinorInterrupt, "Execute skipping a cycle to allow old"
" branch to complete\n");
} else {
ThreadID commit_tid = getCommittingThread();
if (commit_tid != InvalidThreadID) {
ExecuteThreadInfo& commit_info = executeInfo[commit_tid];
DPRINTF(MinorExecute, "Attempting to commit [tid:%d]\n",
commit_tid);
/* commit can set stalled flags observable to issue and so *must* be
* called first */
if (commit_info.drainState != NotDraining) {
if (commit_info.drainState == DrainCurrentInst) {
/* Commit only micro-ops, don't kill anything else */
commit(commit_tid, true, false, branch);
if (isInbetweenInsts(commit_tid))
setDrainState(commit_tid, DrainHaltFetch);
/* Discard any generated branch */
branch = BranchData::bubble();
} else if (commit_info.drainState == DrainAllInsts) {
/* Kill all instructions */
while (getInput(commit_tid))
popInput(commit_tid);
commit(commit_tid, false, true, branch);
}
} else {
/* Commit micro-ops only if interrupted. Otherwise, commit
* anything you like */
DPRINTF(MinorExecute, "Committing micro-ops for interrupt[tid:%d]\n",
commit_tid);
bool only_commit_microops = interrupted &&
hasInterrupt(commit_tid);
commit(commit_tid, only_commit_microops, false, branch);
}
/* Halt fetch, but don't do it until we have the current instruction in
* the bag */
if (commit_info.drainState == DrainHaltFetch) {
updateBranchData(commit_tid, BranchData::HaltFetch,
MinorDynInst::bubble(), TheISA::PCState(0), branch);
cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
setDrainState(commit_tid, DrainAllInsts);
}
}
ThreadID issue_tid = getIssuingThread();
/* This will issue merrily even when interrupted in the sure and
* certain knowledge that the interrupt with change the stream */
if (issue_tid != InvalidThreadID) {
DPRINTF(MinorExecute, "Attempting to issue [tid:%d]\n",
issue_tid);
num_issued = issue(issue_tid);
}
}
/* Run logic to step functional units + decide if we are active on the next
* clock cycle */
std::vector<MinorDynInstPtr> next_issuable_insts;
bool can_issue_next = false;
for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
/* Find the next issuable instruction for each thread and see if it can
be issued */
if (getInput(tid)) {
unsigned int input_index = executeInfo[tid].inputIndex;
MinorDynInstPtr inst = getInput(tid)->insts[input_index];
if (inst->isFault()) {
can_issue_next = true;
} else if (!inst->isBubble()) {
next_issuable_insts.push_back(inst);
}
}
}
bool becoming_stalled = true;
/* Advance the pipelines and note whether they still need to be
* advanced */
for (unsigned int i = 0; i < numFuncUnits; i++) {
FUPipeline *fu = funcUnits[i];
fu->advance();
/* If we need to tick again, the pipeline will have been left or set
* to be unstalled */
if (fu->occupancy !=0 && !fu->stalled)
becoming_stalled = false;
/* Could we possibly issue the next instruction from any thread?
* This is quite an expensive test and is only used to determine
* if the CPU should remain active, only run it if we aren't sure
* we are active next cycle yet */
for (auto inst : next_issuable_insts) {
if (!fu->stalled && fu->provides(inst->staticInst->opClass()) &&
scoreboard[inst->id.threadId].canInstIssue(inst,
NULL, NULL, cpu.curCycle() + Cycles(1),
cpu.getContext(inst->id.threadId))) {
can_issue_next = true;
break;
}
}
}
bool head_inst_might_commit = false;
/* Could the head in flight insts be committed */
for (auto const &info : executeInfo) {
if (!info.inFlightInsts->empty()) {
const QueuedInst &head_inst = info.inFlightInsts->front();
if (head_inst.inst->isNoCostInst()) {
head_inst_might_commit = true;
} else {
FUPipeline *fu = funcUnits[head_inst.inst->fuIndex];
if ((fu->stalled &&
fu->front().inst->id == head_inst.inst->id) ||
lsq.findResponse(head_inst.inst))
{
head_inst_might_commit = true;
break;
}
}
}
}
DPRINTF(Activity, "Need to tick num issued insts: %s%s%s%s%s%s\n",
(num_issued != 0 ? " (issued some insts)" : ""),
(becoming_stalled ? "(becoming stalled)" : "(not becoming stalled)"),
(can_issue_next ? " (can issued next inst)" : ""),
(head_inst_might_commit ? "(head inst might commit)" : ""),
(lsq.needsToTick() ? " (LSQ needs to tick)" : ""),
(interrupted ? " (interrupted)" : ""));
bool need_to_tick =
num_issued != 0 || /* Issued some insts this cycle */
!becoming_stalled || /* Some FU pipelines can still move */
can_issue_next || /* Can still issue a new inst */
head_inst_might_commit || /* Could possible commit the next inst */
lsq.needsToTick() || /* Must step the dcache port */
interrupted; /* There are pending interrupts */
if (!need_to_tick) {
DPRINTF(Activity, "The next cycle might be skippable as there are no"
" advanceable FUs\n");
}
/* Wake up if we need to tick again */
if (need_to_tick)
cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
/* Note activity of following buffer */
if (!branch.isBubble())
cpu.activityRecorder->activity();
/* Make sure the input (if any left) is pushed */
if (!inp.outputWire->isBubble())
inputBuffer[inp.outputWire->threadId].pushTail();
}
ThreadID
Execute::checkInterrupts(BranchData& branch, bool& interrupted)
{
ThreadID tid = interruptPriority;
/* Evaluate interrupts in round-robin based upon service */
do {
/* Has an interrupt been signalled? This may not be acted on
* straighaway so this is different from took_interrupt */
bool thread_interrupted = false;
if (FullSystem && cpu.getInterruptController(tid)) {
/* This is here because it seems that after drainResume the
* interrupt controller isn't always set */
thread_interrupted = executeInfo[tid].drainState == NotDraining &&
isInterrupted(tid);
interrupted = interrupted || thread_interrupted;
} else {
DPRINTF(MinorInterrupt, "No interrupt controller\n");
}
DPRINTF(MinorInterrupt, "[tid:%d] thread_interrupted?=%d isInbetweenInsts?=%d\n",
tid, thread_interrupted, isInbetweenInsts(tid));
/* Act on interrupts */
if (thread_interrupted && isInbetweenInsts(tid)) {
if (takeInterrupt(tid, branch)) {
interruptPriority = tid;
return tid;
}
} else {
tid = (tid + 1) % cpu.numThreads;
}
} while (tid != interruptPriority);
return InvalidThreadID;
}
bool
Execute::hasInterrupt(ThreadID thread_id)
{
if (FullSystem && cpu.getInterruptController(thread_id)) {
return executeInfo[thread_id].drainState == NotDraining &&
isInterrupted(thread_id);
}
return false;
}
void
Execute::minorTrace() const
{
std::ostringstream insts;
std::ostringstream stalled;
executeInfo[0].instsBeingCommitted.reportData(insts);
lsq.minorTrace();
inputBuffer[0].minorTrace();
scoreboard[0].minorTrace();
/* Report functional unit stalling in one string */
unsigned int i = 0;
while (i < numFuncUnits)
{
stalled << (funcUnits[i]->stalled ? '1' : 'E');
i++;
if (i != numFuncUnits)
stalled << ',';
}
MINORTRACE("insts=%s inputIndex=%d streamSeqNum=%d"
" stalled=%s drainState=%d isInbetweenInsts=%d\n",
insts.str(), executeInfo[0].inputIndex, executeInfo[0].streamSeqNum,
stalled.str(), executeInfo[0].drainState, isInbetweenInsts(0));
std::for_each(funcUnits.begin(), funcUnits.end(),
std::mem_fun(&FUPipeline::minorTrace));
executeInfo[0].inFlightInsts->minorTrace();
executeInfo[0].inFUMemInsts->minorTrace();
}
inline ThreadID
Execute::getCommittingThread()
{
std::vector<ThreadID> priority_list;
switch (cpu.threadPolicy) {
case Enums::SingleThreaded:
return 0;
case Enums::RoundRobin:
priority_list = cpu.roundRobinPriority(commitPriority);
break;
case Enums::Random:
priority_list = cpu.randomPriority();
break;
default:
panic("Invalid thread policy");
}
for (auto tid : priority_list) {
ExecuteThreadInfo &ex_info = executeInfo[tid];
bool can_commit_insts = !ex_info.inFlightInsts->empty();
if (can_commit_insts) {
QueuedInst *head_inflight_inst = &(ex_info.inFlightInsts->front());
MinorDynInstPtr inst = head_inflight_inst->inst;
can_commit_insts = can_commit_insts &&
(!inst->inLSQ || (lsq.findResponse(inst) != NULL));
if (!inst->inLSQ) {
bool can_transfer_mem_inst = false;
if (!ex_info.inFUMemInsts->empty() && lsq.canRequest()) {
const MinorDynInstPtr head_mem_ref_inst =
ex_info.inFUMemInsts->front().inst;
FUPipeline *fu = funcUnits[head_mem_ref_inst->fuIndex];
const MinorDynInstPtr &fu_inst = fu->front().inst;
can_transfer_mem_inst =
!fu_inst->isBubble() &&
fu_inst->id.threadId == tid &&
!fu_inst->inLSQ &&
fu_inst->canEarlyIssue &&
inst->id.execSeqNum > fu_inst->instToWaitFor;
}
bool can_execute_fu_inst = inst->fuIndex == noCostFUIndex;
if (can_commit_insts && !can_transfer_mem_inst &&
inst->fuIndex != noCostFUIndex)
{
QueuedInst& fu_inst = funcUnits[inst->fuIndex]->front();
can_execute_fu_inst = !fu_inst.inst->isBubble() &&
fu_inst.inst->id == inst->id;
}
can_commit_insts = can_commit_insts &&
(can_transfer_mem_inst || can_execute_fu_inst);
}
}
if (can_commit_insts) {
commitPriority = tid;
return tid;
}
}
return InvalidThreadID;
}
inline ThreadID
Execute::getIssuingThread()
{
std::vector<ThreadID> priority_list;
switch (cpu.threadPolicy) {
case Enums::SingleThreaded:
return 0;
case Enums::RoundRobin:
priority_list = cpu.roundRobinPriority(issuePriority);
break;
case Enums::Random:
priority_list = cpu.randomPriority();
break;
default:
panic("Invalid thread scheduling policy.");
}
for (auto tid : priority_list) {
if (getInput(tid)) {
issuePriority = tid;
return tid;
}
}
return InvalidThreadID;
}
void
Execute::drainResume()
{
DPRINTF(Drain, "MinorExecute drainResume\n");
for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
setDrainState(tid, NotDraining);
}
cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
}
std::ostream &operator <<(std::ostream &os, Execute::DrainState state)
{
switch (state)
{
case Execute::NotDraining:
os << "NotDraining";
break;
case Execute::DrainCurrentInst:
os << "DrainCurrentInst";
break;
case Execute::DrainHaltFetch:
os << "DrainHaltFetch";
break;
case Execute::DrainAllInsts:
os << "DrainAllInsts";
break;
default:
os << "Drain-" << static_cast<int>(state);
break;
}
return os;
}
void
Execute::setDrainState(ThreadID thread_id, DrainState state)
{
DPRINTF(Drain, "setDrainState[%d]: %s\n", thread_id, state);
executeInfo[thread_id].drainState = state;
}
unsigned int
Execute::drain()
{
DPRINTF(Drain, "MinorExecute drain\n");
for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
if (executeInfo[tid].drainState == NotDraining) {
cpu.wakeupOnEvent(Pipeline::ExecuteStageId);
/* Go to DrainCurrentInst if we're between microops
* or waiting on an unbufferable memory operation.
* Otherwise we can go straight to DrainHaltFetch
*/
if (isInbetweenInsts(tid))
setDrainState(tid, DrainHaltFetch);
else
setDrainState(tid, DrainCurrentInst);
}
}
return (isDrained() ? 0 : 1);
}
bool
Execute::isDrained()
{
if (!lsq.isDrained())
return false;
for (ThreadID tid = 0; tid < cpu.numThreads; tid++) {
if (!inputBuffer[tid].empty() ||
!executeInfo[tid].inFlightInsts->empty()) {
return false;
}
}
return true;
}
Execute::~Execute()
{
for (unsigned int i = 0; i < numFuncUnits; i++)
delete funcUnits[i];
for (ThreadID tid = 0; tid < cpu.numThreads; tid++)
delete executeInfo[tid].inFlightInsts;
}
bool
Execute::instIsRightStream(MinorDynInstPtr inst)
{
return inst->id.streamSeqNum == executeInfo[inst->id.threadId].streamSeqNum;
}
bool
Execute::instIsHeadInst(MinorDynInstPtr inst)
{
bool ret = false;
if (!executeInfo[inst->id.threadId].inFlightInsts->empty())
ret = executeInfo[inst->id.threadId].inFlightInsts->front().inst->id == inst->id;
return ret;
}
MinorCPU::MinorCPUPort &
Execute::getDcachePort()
{
return lsq.getDcachePort();
}
}
Reference in a new issue View git blame Copy permalink