void CoreRecorder::recordAccess(uint64_t startCycle) {
assert(eventRecorder.hasRecord());
TimingRecord tr = eventRecorder.popRecord();
TimingEvent* origPrevResp = prevRespEvent;
assert(startCycle >= prevRespCycle);
assert(tr.reqCycle >= startCycle);
if (IsGet(tr.type)) {
uint64_t delay = tr.reqCycle - prevRespCycle;
TimingEvent* ev = new (eventRecorder) TimingCoreEvent(delay, prevRespCycle - gapCycles, this);
ev->setMinStartCycle(prevRespCycle);
prevRespEvent->addChild(ev, eventRecorder)->addChild(tr.startEvent, eventRecorder);
prevRespEvent = tr.endEvent;
prevRespCycle = tr.respCycle;
assert(prevRespEvent);
} else {
assert(IsPut(tr.type));
// Link previous response and this req directly (don't even create a new event)
DelayEvent* dr = new (eventRecorder) DelayEvent(tr.reqCycle - prevRespCycle);
dr->setMinStartCycle(prevRespCycle);
prevRespEvent->addChild(dr, eventRecorder)->addChild(tr.startEvent, eventRecorder);
//tr.endEvent not linked to anything, it's a PUT
}
origPrevResp->produceCrossings(&eventRecorder);
eventRecorder.getCrossingStack().clear();
}
void TimingEvent::produceCrossings(EventRecorder* evRec) {
assert(domain != -1);
//assert(dynamic_cast<CrossingEvent*>(this) == nullptr); //careful, expensive...
auto pcLambda = [this, evRec](TimingEvent** childPtr) {
TimingEvent* c = *childPtr;
if (c->domain != domain) *childPtr = handleCrossing(c, evRec, true);
c->produceCrossings(evRec);
};
visitChildren< decltype(pcLambda) > (pcLambda);
}
double TimerEx::GetAvgTime( TimerType type, LPCWSTR timerId, bool stall )
{
_ASSERT( "init not called or called with NULL" && (m_pDev != NULL) );
TimingEvent* te = NULL;
if( NULL != m_Current )
te = m_Current->GetTimer( timerId );
if( NULL == te )
te = GetTimer( timerId );
return ( NULL != te ) ? te->GetAvgTime(type, stall) : 0.0;
}
void CoreRecorder::recordAccess(uint64_t startCycle) {
assert(eventRecorder.numRecords() <= 2);
TimingRecord tr = eventRecorder.getRecord(0);
TimingEvent* origPrevResp = prevRespEvent;
if (tr.type == PUTS || tr.type == PUTX) {
//info("Handling PUT+GET");
assert(eventRecorder.numRecords() == 2);
TimingRecord tr1 = eventRecorder.getRecord(1);
assert(tr1.type == GETX || tr1.type == GETS);
assert(startCycle >= prevRespCycle);
assert(tr1.reqCycle >= startCycle);
assert(tr.reqCycle >= startCycle);
uint64_t delay = startCycle - prevRespCycle;
TimingCoreEvent* ev = new (eventRecorder) TimingCoreEvent(delay, prevRespCycle - gapCycles, this);
ev->setMinStartCycle(prevRespCycle);
prevRespEvent->addChild(ev, eventRecorder);
DelayEvent* dr = new (eventRecorder) DelayEvent(tr.reqCycle-startCycle);
DelayEvent* dr1 = new (eventRecorder) DelayEvent(tr1.reqCycle-startCycle);
dr->setMinStartCycle(startCycle);
dr1->setMinStartCycle(startCycle);
ev->addChild(dr, eventRecorder)->addChild(tr.startEvent, eventRecorder);
ev->addChild(dr1, eventRecorder)->addChild(tr1.startEvent, eventRecorder);
//tr.endEvent not linked to anything
prevRespEvent = tr1.endEvent;
prevRespCycle = tr1.respCycle;
} else {
//info("Handling single GET");
assert(tr.type == GETX || tr.type == GETS);
assert(eventRecorder.numRecords() == 1);
uint64_t delay = tr.reqCycle - prevRespCycle;
TimingEvent* ev = new (eventRecorder) TimingCoreEvent(delay, prevRespCycle - gapCycles, this);
ev->setMinStartCycle(prevRespCycle);
prevRespEvent->addChild(ev, eventRecorder)->addChild(tr.startEvent, eventRecorder);
prevRespEvent = tr.endEvent;
prevRespCycle = tr.respCycle;
}
origPrevResp->produceCrossings(&eventRecorder);
eventRecorder.getCrossingStack().clear();
eventRecorder.clearRecords();
}
TimingEvent* TimerEx::GetTimer( LPCWSTR timerId )
{
if( NULL == timerId )
return m_Root;
_ASSERT( "init not called" && (m_pDev != NULL) );
size_t len = wcslen(timerId);
size_t seperator = wcscspn( timerId, L"/|\\" );
if( seperator < len )
{
LPWSTR idCopy = new WCHAR[len+1];
wcscpy_s( idCopy, len+1, timerId );
idCopy[seperator] = 0;
TimingEvent* te = m_Root;
while( te )
{
if( !wcscmp(idCopy, te->m_name ) )
{
te = te->GetTimerRec( &idCopy[seperator+1] );
delete[] idCopy;
return te;
}
te = te->m_next;
}
delete[] idCopy;
}
else
{
TimingEvent* te = m_Root;
while( te )
{
if( !wcscmp(timerId, te->m_name ) )
return te;
te = te->m_next;
}
}
return NULL;
}
// when this function is called we know we're working on a copy of the name, so we can "destruct" it
TimingEvent* TimingEvent::GetTimerRec( LPWSTR timerId )
{
size_t len = wcslen(timerId);
size_t seperator = wcscspn( timerId, L"/|\\" );
if( seperator<len )
timerId[seperator] = 0;
TimingEvent* te = m_firstChild;
while( te )
{
if( !wcscmp(timerId, te->m_name ) )
{
if( seperator<len )
te = te->GetTimerRec( &timerId[seperator+1] );
return te;
}
te = te->m_next;
}
return NULL;
}
TimingEvent* TimingEvent::GetTimer(LPCWSTR timerId)
{
size_t len = wcslen(timerId);
size_t seperator = wcscspn( timerId, L"/|\\" );
if( seperator<len )
{
LPWSTR idCopy = new WCHAR[len+1];
wcscpy_s( idCopy, len+1, timerId );
idCopy[seperator] = 0;
TimingEvent* te = m_firstChild;
while( te )
{
if( !wcscmp(idCopy, te->m_name ) )
{
te = te->GetTimerRec( &idCopy[seperator+1] );
delete[] idCopy;
return te;
}
te = te->m_next;
}
delete[] idCopy;
}
else
{
TimingEvent* te = m_firstChild;
while( te )
{
if( !wcscmp(timerId, te->m_name ) )
return te;
te = te->m_next;
}
}
return NULL;
}
void TimerEx::Start( LPCWSTR timerId )
{
_ASSERT( "init not called or called with NULL" && (m_pDev != NULL) );
TimingEvent* te = (NULL == m_Current) ? GetTimer(timerId) : m_Current->GetTimer(timerId);
if( NULL == te )
{
// create new timer event
if( NULL == m_Unused )
{
te = new TimingEvent();
}
else
{
te = m_Unused;
m_Unused = te->m_next;
te->m_next = NULL;
}
te->SetName( timerId );
te->m_parent = m_Current;
// now look where to insert it
TimingEvent* lu = NULL;
if( NULL == m_Current )
{
TimingEvent* tmp = m_Root;
while( tmp )
{
if( tmp->m_used )
lu = tmp;
tmp = tmp->m_next;
}
}
else
{
lu = m_Current->FindLastChildUsed();
}
if( NULL != lu )
{
te->m_next = lu->m_next;
lu->m_next = te;
}
else
{
if( NULL == m_Current )
{
te->m_next = m_Root;
m_Root = te;
}
else
{
te->m_next = m_Current->m_firstChild;
m_Current->m_firstChild = te;
}
}
}
m_Current = te;
m_Current->Start();
}
// TODO(dsm): This is copied verbatim from Cache. We should split Cache into different methods, then call those.
uint64_t TimingCache::access(MemReq& req) {
EventRecorder* evRec = zinfo->eventRecorders[req.srcId];
assert_msg(evRec, "TimingCache is not connected to TimingCore");
uint32_t initialRecords = evRec->numRecords();
bool hasWritebackRecord = false;
TimingRecord writebackRecord;
bool hasAccessRecord = false;
TimingRecord accessRecord;
uint64_t evDoneCycle = 0;
uint64_t respCycle = req.cycle;
bool skipAccess = cc->startAccess(req); //may need to skip access due to races (NOTE: may change req.type!)
if (likely(!skipAccess)) {
bool updateReplacement = (req.type == GETS) || (req.type == GETX);
int32_t lineId = array->lookup(req.lineAddr, &req, updateReplacement);
respCycle += accLat;
if (lineId == -1 /*&& cc->shouldAllocate(req)*/) {
assert(cc->shouldAllocate(req)); //dsm: for now, we don't deal with non-inclusion in TimingCache
//Make space for new line
Address wbLineAddr;
lineId = array->preinsert(req.lineAddr, &req, &wbLineAddr); //find the lineId to replace
trace(Cache, "[%s] Evicting 0x%lx", name.c_str(), wbLineAddr);
//Evictions are not in the critical path in any sane implementation -- we do not include their delays
//NOTE: We might be "evicting" an invalid line for all we know. Coherence controllers will know what to do
evDoneCycle = cc->processEviction(req, wbLineAddr, lineId, respCycle); //if needed, send invalidates/downgrades to lower level, and wb to upper level
array->postinsert(req.lineAddr, &req, lineId); //do the actual insertion. NOTE: Now we must split insert into a 2-phase thing because cc unlocks us.
if (evRec->numRecords() > initialRecords) {
assert_msg(evRec->numRecords() == initialRecords + 1, "evRec records on eviction %ld", evRec->numRecords());
writebackRecord = evRec->getRecord(initialRecords);
hasWritebackRecord = true;
evRec->popRecord();
}
}
uint64_t getDoneCycle = respCycle;
respCycle = cc->processAccess(req, lineId, respCycle, &getDoneCycle);
if (evRec->numRecords() > initialRecords) {
assert_msg(evRec->numRecords() == initialRecords + 1, "evRec records %ld", evRec->numRecords());
accessRecord = evRec->getRecord(initialRecords);
hasAccessRecord = true;
evRec->popRecord();
}
// At this point we have all the info we need to hammer out the timing record
TimingRecord tr = {req.lineAddr << lineBits, req.cycle, respCycle, req.type, NULL, NULL}; //note the end event is the response, not the wback
if (getDoneCycle - req.cycle == accLat) {
// Hit
assert(!hasWritebackRecord);
assert(!hasAccessRecord);
uint64_t hitLat = respCycle - req.cycle; // accLat + invLat
HitEvent* ev = new (evRec) HitEvent(this, hitLat, domain);
ev->setMinStartCycle(req.cycle);
tr.startEvent = tr.endEvent = ev;
} else {
assert_msg(getDoneCycle == respCycle, "gdc %ld rc %ld", getDoneCycle, respCycle);
// Miss events:
// MissStart (does high-prio lookup) -> getEvent || evictionEvent || replEvent (if needed) -> MissWriteback
MissStartEvent* mse = new (evRec) MissStartEvent(this, accLat, domain);
MissResponseEvent* mre = new (evRec) MissResponseEvent(this, mse, domain);
MissWritebackEvent* mwe = new (evRec) MissWritebackEvent(this, mse, accLat, domain);
mse->setMinStartCycle(req.cycle);
mre->setMinStartCycle(getDoneCycle);
mwe->setMinStartCycle(MAX(evDoneCycle, getDoneCycle));
// Tie two events to an optional timing record
// TODO: Promote to evRec if this is more generally useful
auto connect = [evRec](const TimingRecord* r, TimingEvent* startEv, TimingEvent* endEv, uint64_t startCycle, uint64_t endCycle) {
assert_msg(startCycle <= endCycle, "start > end? %ld %ld", startCycle, endCycle);
if (r) {
assert_msg(startCycle <= r->reqCycle, "%ld / %ld", startCycle, r->reqCycle);
assert_msg(r->respCycle <= endCycle, "%ld %ld %ld %ld", startCycle, r->reqCycle, r->respCycle, endCycle);
uint64_t upLat = r->reqCycle - startCycle;
uint64_t downLat = endCycle - r->respCycle;
if (upLat) {
DelayEvent* dUp = new (evRec) DelayEvent(upLat);
dUp->setMinStartCycle(startCycle);
startEv->addChild(dUp, evRec)->addChild(r->startEvent, evRec);
} else {
startEv->addChild(r->startEvent, evRec);
}
if (downLat) {
DelayEvent* dDown = new (evRec) DelayEvent(downLat);
dDown->setMinStartCycle(r->respCycle);
r->endEvent->addChild(dDown, evRec)->addChild(endEv, evRec);
} else {
r->endEvent->addChild(endEv, evRec);
}
} else {
if (startCycle == endCycle) {
startEv->addChild(endEv, evRec);
} else {
DelayEvent* dEv = new (evRec) DelayEvent(endCycle - startCycle);
dEv->setMinStartCycle(startCycle);
startEv->addChild(dEv, evRec)->addChild(endEv, evRec);
}
}
};
// Get path
connect(hasAccessRecord? &accessRecord : NULL, mse, mre, req.cycle + accLat, getDoneCycle);
mre->addChild(mwe, evRec);
// Eviction path
if (evDoneCycle) {
connect(hasWritebackRecord? &writebackRecord : NULL, mse, mwe, req.cycle + accLat, evDoneCycle);
}
// Replacement path
if (evDoneCycle && cands > ways) {
uint32_t replLookups = (cands + (ways-1))/ways - 1; // e.g., with 4 ways, 5-8 -> 1, 9-12 -> 2, etc.
assert(replLookups);
uint32_t fringeAccs = ways - 1;
uint32_t accsSoFar = 0;
TimingEvent* p = mse;
// Candidate lookup events
while (accsSoFar < replLookups) {
uint32_t preDelay = accsSoFar? 0 : tagLat;
uint32_t postDelay = tagLat - MIN(tagLat - 1, fringeAccs);
uint32_t accs = MIN(fringeAccs, replLookups - accsSoFar);
//info("ReplAccessEvent rl %d fa %d preD %d postD %d accs %d", replLookups, fringeAccs, preDelay, postDelay, accs);
ReplAccessEvent* raEv = new (evRec) ReplAccessEvent(this, accs, preDelay, postDelay, domain);
raEv->setMinStartCycle(req.cycle /*lax...*/);
accsSoFar += accs;
p->addChild(raEv, evRec);
p = raEv;
fringeAccs *= ways - 1;
}
// Swap events -- typically, one read and one write work for 1-2 swaps. Exact number depends on layout.
ReplAccessEvent* rdEv = new (evRec) ReplAccessEvent(this, 1, tagLat, tagLat, domain);
rdEv->setMinStartCycle(req.cycle /*lax...*/);
ReplAccessEvent* wrEv = new (evRec) ReplAccessEvent(this, 1, 0, 0, domain);
wrEv->setMinStartCycle(req.cycle /*lax...*/);
p->addChild(rdEv, evRec)->addChild(wrEv, evRec)->addChild(mwe, evRec);
}
tr.startEvent = mse;
tr.endEvent = mre; // note the end event is the response, not the wback
}
evRec->pushRecord(tr);
}
cc->endAccess(req);
assert_msg(respCycle >= req.cycle, "[%s] resp < req? 0x%lx type %s childState %s, respCycle %ld reqCycle %ld",
name.c_str(), req.lineAddr, AccessTypeName(req.type), MESIStateName(*req.state), respCycle, req.cycle);
return respCycle;
}
