uint64_t Cache::access(MemReq& req) { uint64_t respCycle = req.cycle; bool skipAccess = cc->startAccess(req); //may need to skip access due to races (NOTE: may change req.type!) if( skipAccess) std::cout<<"skip access"<<std::endl; 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)) { //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 cc->processEviction(req, wbLineAddr, lineId, respCycle); //1. if needed, send invalidates/downgrades to lower 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. } respCycle = cc->processAccess(req, lineId, respCycle); } 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; }
int main(int argc, const char* argv[]) { InitLog(""); //no log header if (argc != 2) { info("Prints an access trace"); info("Usage: %s <trace>", argv[0]); exit(1); } gm_init(32<<20 /*32 MB, should be enough*/); AccessTraceReader tr(argv[1]); info("%12s %6s %6s %20s %10s", "Cycle", "Src", "Type", "LineAddr", "Latency"); while(!tr.empty()) { AccessRecord acc = tr.read(); info("%12ld %6d %s %20p %10d", acc.reqCycle, acc.childId, AccessTypeName(acc.type), (uint64_t*)acc.lineAddr, acc.latency); } return 0; }
uint64_t MESIBottomCC::processNonInclusiveWriteback(Address lineAddr, AccessType type, uint64_t cycle, MESIState* state, uint32_t srcId, uint32_t flags) { if (!nonInclusiveHack) panic("Non-inclusive %s on line 0x%lx, this cache should be inclusive", AccessTypeName(type), lineAddr); //info("Non-inclusive wback, forwarding"); MemReq req = {lineAddr, type, selfId, state, cycle, &ccLock, *state, srcId, flags | MemReq::NONINCLWB}; uint64_t respCycle = parents[getParentId(lineAddr)]->access(req); return respCycle; }
// 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; }