diff options
author | Radhika Jagtap <radhika.jagtap@ARM.com> | 2015-12-07 16:42:15 -0600 |
---|---|---|
committer | Radhika Jagtap <radhika.jagtap@ARM.com> | 2015-12-07 16:42:15 -0600 |
commit | de4dc50e9561ab789db01eb4adea1522de623782 (patch) | |
tree | 2cd8c379d8621a78bc87109d107f4d1624222ebd /src/cpu/o3/probe/elastic_trace.cc | |
parent | eb19fc29761a99ff7d5a5e4588866b487a2082ee (diff) | |
download | gem5-de4dc50e9561ab789db01eb4adea1522de623782.tar.xz |
proto, probe: Add elastic trace probe to o3 cpu
The elastic trace is a type of probe listener and listens to probe points
in multiple stages of the O3CPU. The notify method is called on a probe
point typically when an instruction successfully progresses through that
stage.
As different listener methods mapped to the different probe points execute,
relevant information about the instruction, e.g. timestamps and register
accesses, are captured and stored in temporary InstExecInfo class objects.
When the instruction progresses through the commit stage, the timing and the
dependency information about the instruction is finalised and encapsulated in
a struct called TraceInfo. TraceInfo objects are collected in a list instead
of writing them out to the trace file one a time. This is required as the
trace is processed in chunks to evaluate order dependencies and computational
delay in case an instruction does not have any register dependencies. By this
we achieve a simpler algorithm during replay because every record in the
trace can be hooked onto a record in its past. The instruction dependency
trace is written out as a protobuf format file. A second trace containing
fetch requests at absolute timestamps is written to a separate protobuf
format file.
If the instruction is not executed then it is not added to the trace.
The code checks if the instruction had a fault, if it predicated
false and thus previous register values were restored or if it was a
load/store that did not have a request (e.g. when the size of the
request is zero). In all these cases the instruction is set as
executed by the Execute stage and is picked up by the commit probe
listener. But a request is not issued and registers are not written.
So practically, skipping these should not hurt the dependency modelling.
If squashing results in squashing younger instructions, it may happen that
the squash probe discards the inst and removes it from the temporary
store but execute stage deals with the instruction in the next cycle which
results in the execute probe seeing this inst as 'new' inst. A sequence
number of the last processed trace record is used to trap these cases and
not add to the temporary store.
The elastic instruction trace and fetch request trace can be read in and
played back by the TraceCPU.
Diffstat (limited to 'src/cpu/o3/probe/elastic_trace.cc')
-rw-r--r-- | src/cpu/o3/probe/elastic_trace.cc | 939 |
1 files changed, 939 insertions, 0 deletions
diff --git a/src/cpu/o3/probe/elastic_trace.cc b/src/cpu/o3/probe/elastic_trace.cc new file mode 100644 index 000000000..a7a0403f9 --- /dev/null +++ b/src/cpu/o3/probe/elastic_trace.cc @@ -0,0 +1,939 @@ +/* + * Copyright (c) 2013 - 2015 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: Radhika Jagtap + * Andreas Hansson + * Thomas Grass + */ + +#include "cpu/o3/probe/elastic_trace.hh" + +#include "base/callback.hh" +#include "base/output.hh" +#include "base/trace.hh" +#include "cpu/reg_class.hh" +#include "debug/ElasticTrace.hh" +#include "mem/packet.hh" + +ElasticTrace::ElasticTrace(const ElasticTraceParams* params) + : ProbeListenerObject(params), + regEtraceListenersEvent(this), + firstWin(true), + lastClearedSeqNum(0), + depWindowSize(params->depWindowSize), + dataTraceStream(nullptr), + instTraceStream(nullptr), + startTraceInst(params->startTraceInst), + allProbesReg(false) +{ + cpu = dynamic_cast<FullO3CPU<O3CPUImpl>*>(params->manager); + fatal_if(!cpu, "Manager of %s is not of type O3CPU and thus does not "\ + "support dependency tracing.\n", name()); + + fatal_if(depWindowSize == 0, "depWindowSize parameter must be non-zero. "\ + "Recommended size is 3x ROB size in the O3CPU.\n"); + + fatal_if(cpu->numThreads > 1, "numThreads = %i, %s supports tracing for"\ + "single-threaded workload only", cpu->numThreads, name()); + // Initialize the protobuf output stream + fatal_if(params->instFetchTraceFile == "", "Assign instruction fetch "\ + "trace file path to instFetchTraceFile"); + fatal_if(params->dataDepTraceFile == "", "Assign data dependency "\ + "trace file path to dataDepTraceFile"); + std::string filename = simout.resolve(name() + "." + + params->instFetchTraceFile); + instTraceStream = new ProtoOutputStream(filename); + filename = simout.resolve(name() + "." + params->dataDepTraceFile); + dataTraceStream = new ProtoOutputStream(filename); + // Create a protobuf message for the header and write it to the stream + ProtoMessage::PacketHeader inst_pkt_header; + inst_pkt_header.set_obj_id(name()); + inst_pkt_header.set_tick_freq(SimClock::Frequency); + instTraceStream->write(inst_pkt_header); + // Create a protobuf message for the header and write it to + // the stream + ProtoMessage::InstDepRecordHeader data_rec_header; + data_rec_header.set_obj_id(name()); + data_rec_header.set_tick_freq(SimClock::Frequency); + data_rec_header.set_window_size(depWindowSize); + dataTraceStream->write(data_rec_header); + // Register a callback to flush trace records and close the output streams. + Callback* cb = new MakeCallback<ElasticTrace, + &ElasticTrace::flushTraces>(this); + registerExitCallback(cb); +} + +void +ElasticTrace::regProbeListeners() +{ + inform("@%llu: regProbeListeners() called, startTraceInst = %llu", + curTick(), startTraceInst); + if (startTraceInst == 0) { + // If we want to start tracing from the start of the simulation, + // register all elastic trace probes now. + regEtraceListeners(); + } else { + // Schedule an event to register all elastic trace probes when + // specified no. of instructions are committed. + cpu->comInstEventQueue[(ThreadID)0]->schedule(®EtraceListenersEvent, + startTraceInst); + } +} + +void +ElasticTrace::regEtraceListeners() +{ + assert(!allProbesReg); + inform("@%llu: No. of instructions committed = %llu, registering elastic" + " probe listeners", curTick(), cpu->numSimulatedInsts()); + // Create new listeners: provide method to be called upon a notify() for + // each probe point. + listeners.push_back(new ProbeListenerArg<ElasticTrace, RequestPtr>(this, + "FetchRequest", &ElasticTrace::fetchReqTrace)); + listeners.push_back(new ProbeListenerArg<ElasticTrace, DynInstPtr>(this, + "Execute", &ElasticTrace::recordExecTick)); + listeners.push_back(new ProbeListenerArg<ElasticTrace, DynInstPtr>(this, + "ToCommit", &ElasticTrace::recordToCommTick)); + listeners.push_back(new ProbeListenerArg<ElasticTrace, DynInstPtr>(this, + "Rename", &ElasticTrace::updateRegDep)); + listeners.push_back(new ProbeListenerArg<ElasticTrace, SeqNumRegPair>(this, + "SquashInRename", &ElasticTrace::removeRegDepMapEntry)); + listeners.push_back(new ProbeListenerArg<ElasticTrace, DynInstPtr>(this, + "Squash", &ElasticTrace::addSquashedInst)); + listeners.push_back(new ProbeListenerArg<ElasticTrace, DynInstPtr>(this, + "Commit", &ElasticTrace::addCommittedInst)); + allProbesReg = true; +} + +void +ElasticTrace::fetchReqTrace(const RequestPtr &req) +{ + + DPRINTFR(ElasticTrace, "Fetch Req %i,(%lli,%lli,%lli),%i,%i,%lli\n", + (MemCmd::ReadReq), + req->getPC(), req->getVaddr(), req->getPaddr(), + req->getFlags(), req->getSize(), curTick()); + + // Create a protobuf message including the request fields necessary to + // recreate the request in the TraceCPU. + ProtoMessage::Packet inst_fetch_pkt; + inst_fetch_pkt.set_tick(curTick()); + inst_fetch_pkt.set_cmd(MemCmd::ReadReq); + inst_fetch_pkt.set_pc(req->getPC()); + inst_fetch_pkt.set_flags(req->getFlags()); + inst_fetch_pkt.set_addr(req->getPaddr()); + inst_fetch_pkt.set_size(req->getSize()); + // Write the message to the stream. + instTraceStream->write(inst_fetch_pkt); +} + +void +ElasticTrace::recordExecTick(const DynInstPtr &dyn_inst) +{ + + // In a corner case, a retired instruction is propagated backward to the + // IEW instruction queue to handle some side-channel information. But we + // must not process an instruction again. So we test the sequence number + // against the lastClearedSeqNum and skip adding the instruction for such + // corner cases. + if (dyn_inst->seqNum <= lastClearedSeqNum) { + DPRINTFR(ElasticTrace, "[sn:%lli] Ignoring in execute as instruction \ + has already retired (mostly squashed)", dyn_inst->seqNum); + // Do nothing as program has proceeded and this inst has been + // propagated backwards to handle something. + return; + } + + DPRINTFR(ElasticTrace, "[sn:%lli] Execute Tick = %i\n", dyn_inst->seqNum, + curTick()); + // Either the execution info object will already exist if this + // instruction had a register dependency recorded in the rename probe + // listener before entering execute stage or it will not exist and will + // need to be created here. + InstExecInfo* exec_info_ptr; + auto itr_exec_info = tempStore.find(dyn_inst->seqNum); + if (itr_exec_info != tempStore.end()) { + exec_info_ptr = itr_exec_info->second; + } else { + exec_info_ptr = new InstExecInfo; + tempStore[dyn_inst->seqNum] = exec_info_ptr; + } + + exec_info_ptr->executeTick = curTick(); + maxTempStoreSize = std::max(tempStore.size(), + (std::size_t)maxTempStoreSize.value()); +} + +void +ElasticTrace::recordToCommTick(const DynInstPtr &dyn_inst) +{ + // If tracing has just been enabled then the instruction at this stage of + // execution is far enough that we cannot gather info about its past like + // the tick it started execution. Simply return until we see an instruction + // that is found in the tempStore. + auto itr_exec_info = tempStore.find(dyn_inst->seqNum); + if (itr_exec_info == tempStore.end()) { + DPRINTFR(ElasticTrace, "recordToCommTick: [sn:%lli] Not in temp store," + " skipping.\n", dyn_inst->seqNum); + return; + } + + DPRINTFR(ElasticTrace, "[sn:%lli] To Commit Tick = %i\n", dyn_inst->seqNum, + curTick()); + InstExecInfo* exec_info_ptr = itr_exec_info->second; + exec_info_ptr->toCommitTick = curTick(); + +} + +void +ElasticTrace::updateRegDep(const DynInstPtr &dyn_inst) +{ + // Get the sequence number of the instruction + InstSeqNum seq_num = dyn_inst->seqNum; + + assert(dyn_inst->seqNum > lastClearedSeqNum); + + // Since this is the first probe activated in the pipeline, create + // a new execution info object to track this instruction as it + // progresses through the pipeline. + InstExecInfo* exec_info_ptr = new InstExecInfo; + tempStore[seq_num] = exec_info_ptr; + + // Loop through the source registers and look up the dependency map. If + // the source register entry is found in the dependency map, add a + // dependency on the last writer. + int8_t max_regs = dyn_inst->numSrcRegs(); + for (int src_idx = 0; src_idx < max_regs; src_idx++) { + // Get the physical register index of the i'th source register. + PhysRegIndex src_reg = dyn_inst->renamedSrcRegIdx(src_idx); + DPRINTFR(ElasticTrace, "[sn:%lli] Check map for src reg %i\n", seq_num, + src_reg); + auto itr_last_writer = physRegDepMap.find(src_reg); + if (itr_last_writer != physRegDepMap.end()) { + InstSeqNum last_writer = itr_last_writer->second; + // Additionally the dependency distance is kept less than the window + // size parameter to limit the memory allocation to nodes in the + // graph. If the window were tending to infinite we would have to + // load a large number of node objects during replay. + if (seq_num - last_writer < depWindowSize) { + // Record a physical register dependency. + exec_info_ptr->physRegDepSet.insert(last_writer); + } + } + } + + // Loop through the destination registers of this instruction and update + // the physical register dependency map for last writers to registers. + max_regs = dyn_inst->numDestRegs(); + for (int dest_idx = 0; dest_idx < max_regs; dest_idx++) { + // For data dependency tracking the register must be an int, float or + // CC register and not a Misc register. + TheISA::RegIndex dest_reg = dyn_inst->destRegIdx(dest_idx); + if (regIdxToClass(dest_reg) != MiscRegClass) { + // Get the physical register index of the i'th destination register. + dest_reg = dyn_inst->renamedDestRegIdx(dest_idx); + if (dest_reg != TheISA::ZeroReg) { + DPRINTFR(ElasticTrace, "[sn:%lli] Update map for dest reg %i\n", + seq_num, dest_reg); + physRegDepMap[dest_reg] = seq_num; + } + } + } + maxPhysRegDepMapSize = std::max(physRegDepMap.size(), + (std::size_t)maxPhysRegDepMapSize.value()); +} + +void +ElasticTrace::removeRegDepMapEntry(const SeqNumRegPair &inst_reg_pair) +{ + DPRINTFR(ElasticTrace, "Remove Map entry for Reg %i\n", + inst_reg_pair.second); + auto itr_regdep_map = physRegDepMap.find(inst_reg_pair.second); + if (itr_regdep_map != physRegDepMap.end()) + physRegDepMap.erase(itr_regdep_map); +} + +void +ElasticTrace::addSquashedInst(const DynInstPtr &head_inst) +{ + // If the squashed instruction was squashed before being processed by + // execute stage then it will not be in the temporary store. In this case + // do nothing and return. + auto itr_exec_info = tempStore.find(head_inst->seqNum); + if (itr_exec_info == tempStore.end()) + return; + + // If there is a squashed load for which a read request was + // sent before it got squashed then add it to the trace. + DPRINTFR(ElasticTrace, "Attempt to add squashed inst [sn:%lli]\n", + head_inst->seqNum); + // Get pointer to the execution info object corresponding to the inst. + InstExecInfo* exec_info_ptr = itr_exec_info->second; + if (head_inst->isLoad() && exec_info_ptr->executeTick != MaxTick && + exec_info_ptr->toCommitTick != MaxTick && + head_inst->hasRequest() && + head_inst->getFault() == NoFault) { + // Add record to depTrace with commit parameter as false. + addDepTraceRecord(head_inst, exec_info_ptr, false); + } + // As the information contained is no longer needed, remove the execution + // info object from the temporary store. + clearTempStoreUntil(head_inst); +} + +void +ElasticTrace::addCommittedInst(const DynInstPtr &head_inst) +{ + DPRINTFR(ElasticTrace, "Attempt to add committed inst [sn:%lli]\n", + head_inst->seqNum); + + // Add the instruction to the depTrace. + if (!head_inst->isNop()) { + + // If tracing has just been enabled then the instruction at this stage + // of execution is far enough that we cannot gather info about its past + // like the tick it started execution. Simply return until we see an + // instruction that is found in the tempStore. + auto itr_temp_store = tempStore.find(head_inst->seqNum); + if (itr_temp_store == tempStore.end()) { + DPRINTFR(ElasticTrace, "addCommittedInst: [sn:%lli] Not in temp " + "store, skipping.\n", head_inst->seqNum); + return; + } + + // Get pointer to the execution info object corresponding to the inst. + InstExecInfo* exec_info_ptr = itr_temp_store->second; + assert(exec_info_ptr->executeTick != MaxTick); + assert(exec_info_ptr->toCommitTick != MaxTick); + + // Check if the instruction had a fault, if it predicated false and + // thus previous register values were restored or if it was a + // load/store that did not have a request (e.g. when the size of the + // request is zero). In all these cases the instruction is set as + // executed and is picked up by the commit probe listener. But a + // request is not issued and registers are not written. So practically, + // skipping these should not hurt as execution would not stall on them. + // Alternatively, these could be included merely as a compute node in + // the graph. Removing these for now. If correlation accuracy needs to + // be improved in future these can be turned into comp nodes at the + // cost of bigger traces. + if (head_inst->getFault() != NoFault) { + DPRINTF(ElasticTrace, "%s [sn:%lli] has faulted so " + "skip adding it to the trace\n", + (head_inst->isMemRef() ? "Load/store" : "Comp inst."), + head_inst->seqNum); + } else if (head_inst->isMemRef() && !head_inst->hasRequest()) { + DPRINTF(ElasticTrace, "Load/store [sn:%lli] has no request so " + "skip adding it to the trace\n", head_inst->seqNum); + } else if (!head_inst->readPredicate()) { + DPRINTF(ElasticTrace, "%s [sn:%lli] is predicated false so " + "skip adding it to the trace\n", + (head_inst->isMemRef() ? "Load/store" : "Comp inst."), + head_inst->seqNum); + } else { + // Add record to depTrace with commit parameter as true. + addDepTraceRecord(head_inst, exec_info_ptr, true); + } + } + // As the information contained is no longer needed, remove the execution + // info object from the temporary store. + clearTempStoreUntil(head_inst); +} + +void +ElasticTrace::addDepTraceRecord(const DynInstPtr &head_inst, + InstExecInfo* exec_info_ptr, bool commit) +{ + // Create a record to assign dynamic intruction related fields. + TraceInfo* new_record = new TraceInfo; + // Add to map for sequence number look up to retrieve the TraceInfo pointer + traceInfoMap[head_inst->seqNum] = new_record; + + // Assign fields from the instruction + new_record->instNum = head_inst->seqNum; + new_record->load = head_inst->isLoad(); + new_record->store = head_inst->isStore(); + new_record->commit = commit; + + // Assign fields for creating a request in case of a load/store + new_record->reqFlags = head_inst->memReqFlags; + new_record->addr = head_inst->physEffAddrLow; + // Currently the tracing does not support split requests. + new_record->size = head_inst->effSize; + new_record->pc = head_inst->instAddr(); + + // Assign the timing information stored in the execution info object + new_record->executeTick = exec_info_ptr->executeTick; + new_record->toCommitTick = exec_info_ptr->toCommitTick; + new_record->commitTick = curTick(); + + // Assign initial values for number of dependents and computational delay + new_record->numDepts = 0; + new_record->compDelay = -1; + + // The physical register dependency set of the first instruction is + // empty. Since there are no records in the depTrace at this point, the + // case of adding an ROB dependency by using a reverse iterator is not + // applicable. Thus, populate the fields of the record corresponding to the + // first instruction and return. + if (depTrace.empty()) { + // Store the record in depTrace. + depTrace.push_back(new_record); + DPRINTF(ElasticTrace, "Added first inst record %lli to DepTrace.\n", + new_record->instNum); + return; + } + + // Clear register dependencies for squashed loads as they may be dependent + // on squashed instructions and we do not add those to the trace. + if (head_inst->isLoad() && !commit) { + (exec_info_ptr->physRegDepSet).clear(); + } + + // Assign the register dependencies stored in the execution info object + std::set<InstSeqNum>::const_iterator dep_set_it; + for (dep_set_it = (exec_info_ptr->physRegDepSet).begin(); + dep_set_it != (exec_info_ptr->physRegDepSet).end(); + ++dep_set_it) { + auto trace_info_itr = traceInfoMap.find(*dep_set_it); + if (trace_info_itr != traceInfoMap.end()) { + // The register dependency is valid. Assign it and calculate + // computational delay + new_record->physRegDepList.push_back(*dep_set_it); + DPRINTF(ElasticTrace, "Inst %lli has register dependency on " + "%lli\n", new_record->instNum, *dep_set_it); + TraceInfo* reg_dep = trace_info_itr->second; + reg_dep->numDepts++; + compDelayPhysRegDep(reg_dep, new_record); + ++numRegDep; + } else { + // The instruction that this has a register dependency on was + // not added to the trace because of one of the following + // 1. it was an instruction that had a fault + // 2. it was an instruction that was predicated false and + // previous register values were restored + // 3. it was load/store that did not have a request (e.g. when + // the size of the request is zero but this may not be a fault) + // In all these cases the instruction is set as executed and is + // picked up by the commit probe listener. But a request is not + // issued and registers are not written to in these cases. + DPRINTF(ElasticTrace, "Inst %lli has register dependency on " + "%lli is skipped\n",new_record->instNum, *dep_set_it); + } + } + + // Check for and assign an ROB dependency in addition to register + // dependency before adding the record to the trace. + // As stores have to commit in order a store is dependent on the last + // committed load/store. This is recorded in the ROB dependency. + if (head_inst->isStore()) { + // Look up store-after-store order dependency + updateCommitOrderDep(new_record, false); + // Look up store-after-load order dependency + updateCommitOrderDep(new_record, true); + } + + // In case a node is dependency-free or its dependency got discarded + // because it was outside the window, it is marked ready in the ROB at the + // time of issue. A request is sent as soon as possible. To model this, a + // node is assigned an issue order dependency on a committed instruction + // that completed earlier than it. This is done to avoid the problem of + // determining the issue times of such dependency-free nodes during replay + // which could lead to too much parallelism, thinking conservatively. + if (new_record->robDepList.empty() && new_record->physRegDepList.empty()) { + updateIssueOrderDep(new_record); + } + + // Store the record in depTrace. + depTrace.push_back(new_record); + DPRINTF(ElasticTrace, "Added %s inst %lli to DepTrace.\n", + (commit ? "committed" : "squashed"), new_record->instNum); + + // To process the number of records specified by depWindowSize in the + // forward direction, the depTrace must have twice as many records + // to check for dependencies. + if (depTrace.size() == 2 * depWindowSize) { + + DPRINTF(ElasticTrace, "Writing out trace...\n"); + + // Write out the records which have been processed to the trace + // and remove them from the depTrace. + writeDepTrace(depWindowSize); + + // After the first window, writeDepTrace() must check for valid + // compDelay. + firstWin = false; + } +} + +void +ElasticTrace::updateCommitOrderDep(TraceInfo* new_record, + bool find_load_not_store) +{ + assert(new_record->store); + // Iterate in reverse direction to search for the last committed + // load/store that completed earlier than the new record + depTraceRevItr from_itr(depTrace.end()); + depTraceRevItr until_itr(depTrace.begin()); + TraceInfo* past_record = *from_itr; + uint32_t num_go_back = 0; + + // The execution time of this store is when it is sent, that is committed + Tick execute_tick = curTick(); + // Search for store-after-load or store-after-store order dependency + while (num_go_back < depWindowSize && from_itr != until_itr) { + if (find_load_not_store) { + // Check if previous inst is a load completed earlier by comparing + // with execute tick + if (hasLoadCompleted(past_record, execute_tick)) { + // Assign rob dependency and calculate the computational delay + assignRobDep(past_record, new_record); + ++numOrderDepStores; + return; + } + } else { + // Check if previous inst is a store sent earlier by comparing with + // execute tick + if (hasStoreCommitted(past_record, execute_tick)) { + // Assign rob dependency and calculate the computational delay + assignRobDep(past_record, new_record); + ++numOrderDepStores; + return; + } + } + ++from_itr; + past_record = *from_itr; + ++num_go_back; + } +} + +void +ElasticTrace::updateIssueOrderDep(TraceInfo* new_record) +{ + // Interate in reverse direction to search for the last committed + // record that completed earlier than the new record + depTraceRevItr from_itr(depTrace.end()); + depTraceRevItr until_itr(depTrace.begin()); + TraceInfo* past_record = *from_itr; + + uint32_t num_go_back = 0; + Tick execute_tick = 0; + + if (new_record->load) { + // The execution time of a load is when a request is sent + execute_tick = new_record->executeTick; + ++numIssueOrderDepLoads; + } else if (new_record->store) { + // The execution time of a store is when it is sent, i.e. committed + execute_tick = curTick(); + ++numIssueOrderDepStores; + } else { + // The execution time of a non load/store is when it completes + execute_tick = new_record->toCommitTick; + ++numIssueOrderDepOther; + } + + // We search if this record has an issue order dependency on a past record. + // Once we find it, we update both the new record and the record it depends + // on and return. + while (num_go_back < depWindowSize && from_itr != until_itr) { + // Check if a previous inst is a load sent earlier, or a store sent + // earlier, or a comp inst completed earlier by comparing with execute + // tick + if (hasLoadBeenSent(past_record, execute_tick) || + hasStoreCommitted(past_record, execute_tick) || + hasCompCompleted(past_record, execute_tick)) { + // Assign rob dependency and calculate the computational delay + assignRobDep(past_record, new_record); + return; + } + ++from_itr; + past_record = *from_itr; + ++num_go_back; + } +} + +void +ElasticTrace::assignRobDep(TraceInfo* past_record, TraceInfo* new_record) { + DPRINTF(ElasticTrace, "%s %lli has ROB dependency on %lli\n", + new_record->load ? "Load" : (new_record->store ? "Store" : + "Non load/store"), + new_record->instNum, past_record->instNum); + + // Add dependency on past record + new_record->robDepList.push_back(past_record->instNum); + // Update new_record's compute delay with respect to the past record + compDelayRob(past_record, new_record); + // Increment number of dependents of the past record + ++(past_record->numDepts); + // Update stat to log max number of dependents + maxNumDependents = std::max(past_record->numDepts, + (uint32_t)maxNumDependents.value()); +} + +bool +ElasticTrace::hasStoreCommitted(TraceInfo* past_record, + Tick execute_tick) const +{ + return (past_record->store && past_record->commitTick <= execute_tick); +} + +bool +ElasticTrace::hasLoadCompleted(TraceInfo* past_record, + Tick execute_tick) const +{ + return(past_record->load && past_record->commit && + past_record->toCommitTick <= execute_tick); +} + +bool +ElasticTrace::hasLoadBeenSent(TraceInfo* past_record, + Tick execute_tick) const +{ + // Check if previous inst is a load sent earlier than this + return (past_record->load && past_record->commit && + past_record->executeTick <= execute_tick); +} + +bool +ElasticTrace::hasCompCompleted(TraceInfo* past_record, + Tick execute_tick) const +{ + return(!past_record->store && !past_record->load && + past_record->toCommitTick <= execute_tick); +} + +void +ElasticTrace::clearTempStoreUntil(const DynInstPtr head_inst) +{ + // Clear from temp store starting with the execution info object + // corresponding the head_inst and continue clearing by decrementing the + // sequence number until the last cleared sequence number. + InstSeqNum temp_sn = (head_inst->seqNum); + while (temp_sn > lastClearedSeqNum) { + auto itr_exec_info = tempStore.find(temp_sn); + if (itr_exec_info != tempStore.end()) { + InstExecInfo* exec_info_ptr = itr_exec_info->second; + // Free allocated memory for the info object + delete exec_info_ptr; + // Remove entry from temporary store + tempStore.erase(itr_exec_info); + } + temp_sn--; + } + // Update the last cleared sequence number to that of the head_inst + lastClearedSeqNum = head_inst->seqNum; +} + +void +ElasticTrace::compDelayRob(TraceInfo* past_record, TraceInfo* new_record) +{ + // The computation delay is the delay between the completion tick of the + // inst. pointed to by past_record and the execution tick of its dependent + // inst. pointed to by new_record. + int64_t comp_delay = -1; + Tick execution_tick = 0, completion_tick = 0; + + DPRINTF(ElasticTrace, "Seq num %lli has ROB dependency on seq num %lli.\n", + new_record->instNum, past_record->instNum); + + // Get the tick when the node is executed as per the modelling of + // computation delay + execution_tick = new_record->getExecuteTick(); + + if (past_record->load) { + if (new_record->store) { + completion_tick = past_record->toCommitTick; + } else { + completion_tick = past_record->executeTick; + } + } else if (past_record->store) { + completion_tick = past_record->commitTick; + } else { + completion_tick = past_record->toCommitTick; + } + assert(execution_tick >= completion_tick); + comp_delay = execution_tick - completion_tick; + + DPRINTF(ElasticTrace, "Computational delay is %lli - %lli = %lli\n", + execution_tick, completion_tick, comp_delay); + + // Assign the computational delay with respect to the dependency which + // completes the latest. + if (new_record->compDelay == -1) + new_record->compDelay = comp_delay; + else + new_record->compDelay = std::min(comp_delay, new_record->compDelay); + DPRINTF(ElasticTrace, "Final computational delay = %lli.\n", + new_record->compDelay); +} + +void +ElasticTrace::compDelayPhysRegDep(TraceInfo* past_record, + TraceInfo* new_record) +{ + // The computation delay is the delay between the completion tick of the + // inst. pointed to by past_record and the execution tick of its dependent + // inst. pointed to by new_record. + int64_t comp_delay = -1; + Tick execution_tick = 0, completion_tick = 0; + + DPRINTF(ElasticTrace, "Seq. num %lli has register dependency on seq. num" + " %lli.\n", new_record->instNum, past_record->instNum); + + // Get the tick when the node is executed as per the modelling of + // computation delay + execution_tick = new_record->getExecuteTick(); + + // When there is a physical register dependency on an instruction, the + // completion tick of that instruction is when it wrote to the register, + // that is toCommitTick. In case, of a store updating a destination + // register, this is approximated to commitTick instead + if (past_record->store) { + completion_tick = past_record->commitTick; + } else { + completion_tick = past_record->toCommitTick; + } + assert(execution_tick >= completion_tick); + comp_delay = execution_tick - completion_tick; + DPRINTF(ElasticTrace, "Computational delay is %lli - %lli = %lli\n", + execution_tick, completion_tick, comp_delay); + + // Assign the computational delay with respect to the dependency which + // completes the latest. + if (new_record->compDelay == -1) + new_record->compDelay = comp_delay; + else + new_record->compDelay = std::min(comp_delay, new_record->compDelay); + DPRINTF(ElasticTrace, "Final computational delay = %lli.\n", + new_record->compDelay); +} + +Tick +ElasticTrace::TraceInfo::getExecuteTick() const +{ + if (load) { + // Execution tick for a load instruction is when the request was sent, + // that is executeTick. + return executeTick; + } else if (store) { + // Execution tick for a store instruction is when the request was sent, + // that is commitTick. + return commitTick; + } else { + // Execution tick for a non load/store instruction is when the register + // value was written to, that is commitTick. + return toCommitTick; + } +} + +void +ElasticTrace::writeDepTrace(uint32_t num_to_write) +{ + // Write the trace with fields as follows: + // Instruction sequence number + // If instruction was a load + // If instruction was a store + // If instruction has addr + // If instruction has size + // If instruction has flags + // List of order dependencies - optional, repeated + // Computational delay with respect to last completed dependency + // List of physical register RAW dependencies - optional, repeated + // Weight of a node equal to no. of filtered nodes before it - optional + uint16_t num_filtered_nodes = 0; + depTraceItr dep_trace_itr(depTrace.begin()); + depTraceItr dep_trace_itr_start = dep_trace_itr; + while (num_to_write > 0) { + TraceInfo* temp_ptr = *dep_trace_itr; + // If no node dependends on a non load/store node then there is + // no reason to track it in the dependency graph. We filter out such + // nodes but count them and add a weight field to the subsequent node + // that we do include in the trace. + if (temp_ptr->numDepts != 0 || temp_ptr->load || temp_ptr->store) { + + DPRINTFR(ElasticTrace, "Instruction with seq. num %lli " + "is as follows:\n", temp_ptr->instNum); + if (temp_ptr->load || temp_ptr->store) { + DPRINTFR(ElasticTrace, "\tis a %s\n", + (temp_ptr->load ? "Load" : "Store")); + DPRINTFR(ElasticTrace, "\thas a request with addr %i, size %i," + " flags %i\n", temp_ptr->addr, temp_ptr->size, + temp_ptr->reqFlags); + } else { + DPRINTFR(ElasticTrace, "\tis not a load or store\n"); + } + if (firstWin && temp_ptr->compDelay == -1) { + if (temp_ptr->load) { + temp_ptr->compDelay = temp_ptr->executeTick; + } else if (temp_ptr->store) { + temp_ptr->compDelay = temp_ptr->commitTick; + } else { + temp_ptr->compDelay = temp_ptr->toCommitTick; + } + } + assert(temp_ptr->compDelay != -1); + DPRINTFR(ElasticTrace, "\thas computational delay %lli\n", + temp_ptr->compDelay); + + // Create a protobuf message for the dependency record + ProtoMessage::InstDepRecord dep_pkt; + dep_pkt.set_seq_num(temp_ptr->instNum); + dep_pkt.set_load(temp_ptr->load); + dep_pkt.set_store(temp_ptr->store); + dep_pkt.set_pc(temp_ptr->pc); + if (temp_ptr->load || temp_ptr->store) { + dep_pkt.set_flags(temp_ptr->reqFlags); + dep_pkt.set_addr(temp_ptr->addr); + dep_pkt.set_size(temp_ptr->size); + } + dep_pkt.set_comp_delay(temp_ptr->compDelay); + if (temp_ptr->robDepList.empty()) { + DPRINTFR(ElasticTrace, "\thas no order (rob) dependencies\n"); + } + while (!temp_ptr->robDepList.empty()) { + DPRINTFR(ElasticTrace, "\thas order (rob) dependency on %lli\n", + temp_ptr->robDepList.front()); + dep_pkt.add_rob_dep(temp_ptr->robDepList.front()); + temp_ptr->robDepList.pop_front(); + } + if (temp_ptr->physRegDepList.empty()) { + DPRINTFR(ElasticTrace, "\thas no register dependencies\n"); + } + while (!temp_ptr->physRegDepList.empty()) { + DPRINTFR(ElasticTrace, "\thas register dependency on %lli\n", + temp_ptr->physRegDepList.front()); + dep_pkt.add_reg_dep(temp_ptr->physRegDepList.front()); + temp_ptr->physRegDepList.pop_front(); + } + if (num_filtered_nodes != 0) { + // Set the weight of this node as the no. of filtered nodes + // between this node and the last node that we wrote to output + // stream. The weight will be used during replay to model ROB + // occupancy of filtered nodes. + dep_pkt.set_weight(num_filtered_nodes); + num_filtered_nodes = 0; + } + // Write the message to the protobuf output stream + dataTraceStream->write(dep_pkt); + } else { + // Don't write the node to the trace but note that we have filtered + // out a node. + ++numFilteredNodes; + ++num_filtered_nodes; + } + dep_trace_itr++; + traceInfoMap.erase(temp_ptr->instNum); + delete temp_ptr; + num_to_write--; + } + depTrace.erase(dep_trace_itr_start, dep_trace_itr); +} + +void +ElasticTrace::regStats() { + using namespace Stats; + numRegDep + .name(name() + ".numRegDep") + .desc("Number of register dependencies recorded during tracing") + ; + + numOrderDepStores + .name(name() + ".numOrderDepStores") + .desc("Number of commit order (rob) dependencies for a store recorded" + " on a past load/store during tracing") + ; + + numIssueOrderDepLoads + .name(name() + ".numIssueOrderDepLoads") + .desc("Number of loads that got assigned issue order dependency" + " because they were dependency-free") + ; + + numIssueOrderDepStores + .name(name() + ".numIssueOrderDepStores") + .desc("Number of stores that got assigned issue order dependency" + " because they were dependency-free") + ; + + numIssueOrderDepOther + .name(name() + ".numIssueOrderDepOther") + .desc("Number of non load/store insts that got assigned issue order" + " dependency because they were dependency-free") + ; + + numFilteredNodes + .name(name() + ".numFilteredNodes") + .desc("No. of nodes filtered out before writing the output trace") + ; + + maxNumDependents + .name(name() + ".maxNumDependents") + .desc("Maximum number or dependents on any instruction") + ; + + maxTempStoreSize + .name(name() + ".maxTempStoreSize") + .desc("Maximum size of the temporary store during the run") + ; + + maxPhysRegDepMapSize + .name(name() + ".maxPhysRegDepMapSize") + .desc("Maximum size of register dependency map") + ; +} + +const std::string +ElasticTrace::name() const +{ + return ProbeListenerObject::name(); +} + +void +ElasticTrace::flushTraces() +{ + // Write to trace all records in the depTrace. + writeDepTrace(depTrace.size()); + // Delete the stream objects + delete dataTraceStream; + delete instTraceStream; +} + +ElasticTrace* +ElasticTraceParams::create() +{ + return new ElasticTrace(this); +} |