219 lines
6.0 KiB
C++
219 lines
6.0 KiB
C++
#include "SmartAlgorithm.hpp"
|
|
|
|
#include <algorithm>
|
|
#include <limits>
|
|
|
|
#include "ResultCollector.hpp"
|
|
|
|
extern bool g_use_weights;
|
|
extern bool g_use_watchpoints;
|
|
extern bool g_use_checkpoints;
|
|
|
|
extern unsigned int g_cp_thresh;
|
|
extern unsigned int g_cost_cp;
|
|
|
|
// Must be smaller than ((COST_CP_THRESH - COST_CP) / 2)
|
|
// If too high: Costs per dynamic instruction will approach
|
|
// COST_CP_THRESH but number of CPs can be reduced drastically
|
|
extern unsigned int g_rollback_thresh;
|
|
|
|
extern fail::Logger LOG;
|
|
|
|
bool SmartAlgorithm::calculateAllHops(TraceReader& trace)
|
|
{
|
|
unsigned int trace_pos = 0;
|
|
unsigned int costs = 0;
|
|
|
|
std::vector<trace_event_tuple_t > trace_events;
|
|
std::vector<result_tuple > result;
|
|
|
|
while (trace.getNextTraceEvents(trace_pos, trace_events)) {
|
|
// Policy when multiple trace events: Choose the one with smallest last_pos
|
|
trace_event_tuple_t *best_te = NULL;
|
|
trace_pos_t last_pos = std::numeric_limits<trace_pos_t>::max();
|
|
|
|
bool hop_found = false;
|
|
bool checkpoint_forbidden = false;
|
|
|
|
for (std::vector<trace_event_tuple_t >::iterator it_te = trace_events.begin();
|
|
it_te != trace_events.end();
|
|
it_te++) {
|
|
|
|
// Don't use watchpoints?
|
|
if (!g_use_watchpoints && it_te->second != ACCESS_NONE) {
|
|
continue;
|
|
}
|
|
|
|
// Find last pos of current trace event
|
|
std::map<trace_event_tuple_t, trace_pos_t>::iterator it_lp = m_last_positions.find(*it_te);
|
|
|
|
// Not known? => Single hop, result calculation complete
|
|
if (it_lp == m_last_positions.end()) {
|
|
m_last_positions.insert(std::pair<trace_event_tuple_t, trace_pos_t>(*it_te, trace_pos));
|
|
// New instruction => single hop
|
|
if (!hop_found) {
|
|
result.clear();
|
|
result.push_back(result_tuple(*it_te, trace_pos));
|
|
costs = COST_CHANGE;
|
|
m_resultCollector->addResult(result, costs);
|
|
hop_found = true;
|
|
}
|
|
} else {
|
|
if (it_lp->second < last_pos) {
|
|
last_pos = it_lp->second;
|
|
best_te = (trace_event_tuple_t*)&(it_lp->first);
|
|
}
|
|
it_lp->second = trace_pos;
|
|
}
|
|
}
|
|
|
|
if (hop_found) {
|
|
continue;
|
|
}
|
|
|
|
// Deletion of unnecessary hops
|
|
|
|
/*
|
|
* |----------------|
|
|
* |pos bigger than |
|
|
* |last pos of new | --> delete -->-|
|
|
* |trace_event? | |
|
|
* |----------------| |
|
|
* ^ |
|
|
* | v
|
|
* A B A C <<< Last result
|
|
* ^ ^
|
|
* Reverse Reverse
|
|
* iterator iterator
|
|
* rit_pre_last rit_last
|
|
*
|
|
*/
|
|
|
|
std::vector<result_tuple >::reverse_iterator rit_pre_last, rit_last, end;
|
|
|
|
|
|
rit_last = result.rbegin();
|
|
rit_pre_last = result.rbegin();
|
|
rit_pre_last++;
|
|
|
|
while (rit_last != result.rend()) {
|
|
|
|
trace_pos_t left_pos;
|
|
|
|
// Policy switch:
|
|
// No weights => last pos <= pos of rit_pre_last
|
|
// Weights => last pos < pos of rit_pre_last
|
|
|
|
if ((rit_pre_last == result.rend())) {
|
|
// First node in hop list is Checkpoint? => Look at result before CP-Creation
|
|
if (rit_last->first.second == ACCESS_CHECKPOINT) {
|
|
rit_pre_last = (m_checkpoints[rit_last->first.first]).second.rbegin();
|
|
|
|
// This should never happen:
|
|
if (rit_pre_last == (m_checkpoints[rit_last->first.first]).second.rend()) {
|
|
break;
|
|
}
|
|
|
|
// Don't allow deletion of CP, if afterwards
|
|
// too few of the olf hops would be deleted.
|
|
// (past before to be deleted CP)
|
|
|
|
// Should not exceed vector boundaries, if
|
|
// g_rollback_thresh*2 < (g_cp_thresh - g_cost_cp)
|
|
if ((*(rit_pre_last + g_rollback_thresh)).second < last_pos) {
|
|
checkpoint_forbidden = true;
|
|
break;
|
|
}
|
|
} else {
|
|
break;
|
|
}
|
|
}
|
|
|
|
left_pos = rit_pre_last->second;
|
|
|
|
if ((!g_use_weights && !(last_pos <= left_pos))
|
|
|| (g_use_weights && !(last_pos < left_pos))) {
|
|
break;
|
|
}
|
|
|
|
// Hop a->b is not allowed if last pos of b is at a
|
|
// but a is not b
|
|
// e.g. instruction x with mem access at y
|
|
// a is WP y, and b is BP x
|
|
// OR a is BP x, and b is WP y but instruction
|
|
// at position a has also memory access at y
|
|
if (rit_pre_last != result.rend()) {
|
|
if ((last_pos == rit_pre_last->second) &&
|
|
rit_pre_last->first != *best_te) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
// Right hop will be deleted
|
|
// => Recalculate costs
|
|
|
|
// If to be deleted hop is a checkpoint:
|
|
// There will be no past in the trace
|
|
// Reconfigure result and iterators
|
|
if (rit_last->first.second == ACCESS_CHECKPOINT) {
|
|
costs = m_checkpoints[rit_last->first.first].first;
|
|
|
|
result.clear();
|
|
std::vector<result_tuple > *tmp = &(m_checkpoints[rit_last->first.first].second);
|
|
result.insert(result.end(), tmp->begin(), tmp->end());
|
|
|
|
rit_last = result.rbegin();
|
|
rit_pre_last = result.rbegin();
|
|
rit_pre_last++;
|
|
|
|
// CP could be removed from m_checkpoints, but to do this
|
|
// m_checkpoints needs to be a map (CP-ID is implicit as vector index)
|
|
|
|
continue;
|
|
}
|
|
|
|
if (rit_pre_last->first == rit_last->first) {
|
|
costs -= COST_NO_CHANGE;
|
|
} else {
|
|
costs -= COST_CHANGE;
|
|
}
|
|
|
|
// Delete element described by rit_last
|
|
// Deletion of elements described by reverse iterator works like this
|
|
result.erase((rit_last+1).base());
|
|
rit_last = rit_pre_last;
|
|
rit_pre_last++;
|
|
}
|
|
|
|
|
|
// Add costs for new hop
|
|
if (*best_te == (result.back()).first) {
|
|
costs += COST_NO_CHANGE;
|
|
} else {
|
|
costs += COST_CHANGE;
|
|
}
|
|
|
|
// Check if Checkpoint needed
|
|
if (g_use_checkpoints && (costs > g_cp_thresh) && !checkpoint_forbidden) {
|
|
checkpoint_tuple_t new_cp(costs, std::vector<result_tuple >(result));
|
|
m_checkpoints.push_back(new_cp);
|
|
result.clear();
|
|
result.push_back(result_tuple(trace_event_tuple_t(m_next_cp_id++,
|
|
ACCESS_CHECKPOINT),
|
|
trace_pos));
|
|
costs = g_cost_cp;
|
|
|
|
m_resultCollector->addCheckpoint(trace_pos);
|
|
} else {
|
|
// Add new hop
|
|
result.push_back(result_tuple(*best_te, trace_pos));
|
|
}
|
|
|
|
m_resultCollector->addResult(result, costs);
|
|
}
|
|
// MAX MEM USAGE
|
|
m_resultCollector->setMaxMemUsage();
|
|
|
|
return true;
|
|
}
|