thread speed

dcalc/GraphDelayCalc1.cc

@@ -359,6 +359,7 @@ class FindVertexDelays : public VertexVisitor
   virtual ~FindVertexDelays();
   virtual void visit(Vertex *vertex);
   virtual VertexVisitor *copy();
+  virtual void levelFinished();
 
 protected:
   GraphDelayCalc1 *graph_delay_calc1_;
@@ -396,6 +397,24 @@ FindVertexDelays::visit(Vertex *vertex)
   graph_delay_calc1_->findVertexDelay(vertex, arc_delay_calc_, true);
 }
 
+void
+FindVertexDelays::levelFinished()
+{
+  graph_delay_calc1_->mergeIdealClks();
+}
+
+void
+GraphDelayCalc1::mergeIdealClks()
+{
+  for (auto vertex_clks : ideal_clks_map_next_) {
+    const Vertex *vertex = vertex_clks.first;
+    ClockSet *prev_clks = ideal_clks_map_.findKey(vertex);
+    delete prev_clks;
+    ideal_clks_map_[vertex] = vertex_clks.second;
+  }
+  ideal_clks_map_next_.clear();
+}
+
 // The logical structure of incremental delay calculation closely
 // resembles the incremental search arrival time algorithm
 // (Search::findArrivals).
@@ -417,16 +436,14 @@ GraphDelayCalc1::findDelays(Level level)
     if (incremental_)
       seedInvalidDelays();
 
+    mergeIdealClks();
     FindVertexDelays visitor(this, arc_delay_calc_, false);
     dcalc_count += iter_->visitParallel(level, &visitor);
 
     // Timing checks require slews at both ends of the arc,
     // so find their delays after all slews are known.
-    VertexSet::Iterator check_iter(invalid_checks_);
-    while (check_iter.hasNext()) {
-      Vertex *check_vertex = check_iter.next();
+    for (Vertex *check_vertex : invalid_checks_)
       findCheckDelays(check_vertex, arc_delay_calc_);
-    }
     invalid_checks_.clear();
 
     delays_exist_ = true;
@@ -1628,20 +1645,21 @@ void
 GraphDelayCalc1::clearIdealClkMap()
 {
   ideal_clks_map_.deleteContentsClear();
+  ideal_clks_map_next_.deleteContentsClear();
 }
 
 bool
 GraphDelayCalc1::setIdealClks(const Vertex *vertex,
 			      ClockSet *clks)
 {
   bool changed = false;
-  UniqueLock lock(ideal_clks_map_lock_);
   ClockSet *clks1 = ideal_clks_map_.findKey(vertex);
-  if (ClockSet::equal(clks, clks1))
-    delete clks;
-  else {
-    delete clks1;
-    ideal_clks_map_[vertex] = clks;
+  if (!ClockSet::equal(clks, clks1)) {
+    // Only lock for updates to vertex ideal clks.
+    // Finding ideal clks by level means only changes at the current
+    // delay calc level are changed.
+    UniqueLock lock(ideal_clks_map_next_lock_);
+    ideal_clks_map_next_[vertex] = clks;
     changed = true;
   }
   return changed;
@@ -1650,9 +1668,7 @@ GraphDelayCalc1::setIdealClks(const Vertex *vertex,
 ClockSet *
 GraphDelayCalc1::idealClks(const Vertex *vertex)
 {
-  UniqueLock lock(ideal_clks_map_lock_);
-  ClockSet *clks = ideal_clks_map_.findKey(vertex);
-  return clks;
+  return ideal_clks_map_.findKey(vertex);
 }
 
 bool

dcalc/GraphDelayCalc1.hh

@@ -213,6 +213,7 @@ protected:
 		Parasitic *drvr_parasitic,
 		const TransRiseFall *tr,
 		const DcalcAnalysisPt *dcalc_ap) const;
+  void mergeIdealClks();
 
   // Observer for edge delay changes.
   DelayCalcObserver *observer_;
@@ -234,7 +235,8 @@ protected:
   // delays to be recomputed during incremental delay calculation.
   float incremental_delay_tolerance_;
   VertexIdealClksMap ideal_clks_map_;
-  std::mutex ideal_clks_map_lock_;
+  VertexIdealClksMap ideal_clks_map_next_;
+  std::mutex ideal_clks_map_next_lock_;
 
   friend class FindVertexDelays;
   friend class MultiDrvrNet;

sdc/Sdc.cc

@@ -2502,13 +2502,17 @@ Sdc::cycleAccting(const ClockEdge *src,
   CycleAccting probe(src, tgt);
   CycleAccting *acct = cycle_acctings_.findKey(&probe);
   if (acct == nullptr) {
-    acct = new CycleAccting(src, tgt);
-    if (src == defaultArrivalClockEdge())
-      acct->findDefaultArrivalSrcDelays();
-    else
-      acct->findDelays(this);
     UniqueLock lock(cycle_acctings_lock_);
-    cycle_acctings_.insert(acct);
+    // Recheck with lock.
+    acct = cycle_acctings_.findKey(&probe);
+    if (acct == nullptr) {
+      acct = new CycleAccting(src, tgt);
+      if (src == defaultArrivalClockEdge())
+	acct->findDefaultArrivalSrcDelays();
+      else
+	acct->findDelays(this);
+      cycle_acctings_.insert(acct);
+    }
   }
   return acct;
 }
@@ -2520,7 +2524,9 @@ Sdc::reportClkToClkMaxCycleWarnings()
   // duplicate warnings between different src/tgt clk edges.
   ClockPairSet clk_warnings;
   ClockPairSeq clk_warnings2;
-  for (auto acct : cycle_acctings_) {
+  CycleAcctingSet::Iterator acct_iter(cycle_acctings_);
+  while (acct_iter.hasNext()) {
+    CycleAccting *acct = acct_iter.next();
     if (acct->maxCyclesExceeded()) {
       Clock *src = acct->src()->clock();
       Clock *tgt = acct->target()->clock();
@@ -3301,22 +3307,17 @@ Sdc::setNetWireCap(Net *net,
   }
   if (net_wire_cap_map_ == nullptr)
     net_wire_cap_map_ = new NetWireCapMap[corners_->count()];
-  bool make_drvr_entry = net_wire_cap_map_[corner->index()].hasKey(net);
+  bool make_drvr_entry = !net_wire_cap_map_[corner->index()].hasKey(net);
   MinMaxFloatValues &values = net_wire_cap_map_[corner->index()][net];
   values.setValue(min_max, wire_cap);
 
   // Only need to do this when there is new net_wire_cap_map_ entry.
   if (make_drvr_entry) {
-    NetConnectedPinIterator *pin_iter = network_->connectedPinIterator(net);
-    while (pin_iter->hasNext()) {
-      Pin *pin = pin_iter->next();
-      if (network_->isDriver(pin)) {
-	if (drvr_pin_wire_cap_map_ == nullptr)
-	  drvr_pin_wire_cap_map_ = new PinWireCapMap[corners_->count()];
-	drvr_pin_wire_cap_map_[corner->index()][pin] = &values;
-      }
+    for (Pin *pin : *network_->drivers(net)) {
+      if (drvr_pin_wire_cap_map_ == nullptr)
+	drvr_pin_wire_cap_map_ = new PinWireCapMap[corners_->count()];
+      drvr_pin_wire_cap_map_[corner->index()][pin] = &values;
     }
-    delete pin_iter;
   }
 }
 

sdc/Sdc.hh

@@ -22,7 +22,7 @@
 #include "StringUtil.hh"
 #include "StringSet.hh"
 #include "Map.hh"
-#include "UnorderedSet.hh"
+#include "HashSet.hh"
 #include "UnorderedMap.hh"
 #include "MinMax.hh"
 #include "RiseFallValues.hh"
@@ -114,7 +114,8 @@ typedef Set<InputDelay*> InputDelaySet;
 typedef Map<const Pin*,InputDelaySet*> InputDelayRefPinMap;
 typedef Map<const Pin*,InputDelaySet*> InputDelayInternalPinMap;
 typedef Map<const Pin*,OutputDelay*> OutputDelayMap;
-typedef UnorderedSet<CycleAccting*, CycleAcctingHash, CycleAcctingEqual> CycleAcctingSet;
+// Use HashSet so no read lock is required.
+typedef HashSet<CycleAccting*, CycleAcctingHash, CycleAcctingEqual> CycleAcctingSet;
 typedef Set<Instance*> InstanceSet;
 typedef UnorderedMap<const Pin*,ExceptionPathSet*> PinExceptionsMap;
 typedef Map<const Clock*,ExceptionPathSet*> ClockExceptionsMap;

search/Bfs.cc

@@ -153,6 +153,7 @@ BfsIterator::visit(Level to_level,
 	}
       }
       level_vertices.clear();
+      visitor->levelFinished();
     }
   }
   return visit_count;
@@ -224,7 +225,7 @@ BfsIterator::visitParallel(Level to_level,
 
 	  for (int i = 0; i < thread_count_; i++) {
 	    ForEachArg<QueueIterator, VertexVisitor> arg(&iter, lock,
-							 visitor->copy());
+						      visitor->copy());
 	    // Missing check for null vertex.
 	    threads.push_back(std::thread(forEachBegin<QueueIterator,
 					  VertexVisitor, Vertex*>, arg));
@@ -235,6 +236,7 @@ BfsIterator::visitParallel(Level to_level,
 	    thread.join();
 
 	  visit_count += iter.count();
+	  visitor->levelFinished();
 	  level = first_level_;
 	}
 	else {

search/Sim.cc

@@ -1079,12 +1079,9 @@ Sim::annotateVertexEdges(const Instance *inst,
   InstancePinIterator *pin_iter = network_->pinIterator(inst);
   while (pin_iter->hasNext()) {
     Pin *pin = pin_iter->next();
-    Vertex *vertex, *bidirect_drvr_vertex;
-    graph_->pinVertices(pin, vertex, bidirect_drvr_vertex);
+    Vertex *vertex = graph_->pinDrvrVertex(pin);
     if (vertex)
       annotateVertexEdges(inst, pin, vertex, annotate);
-    if (bidirect_drvr_vertex)
-      annotateVertexEdges(inst, pin, bidirect_drvr_vertex, annotate);
   }
   delete pin_iter;
 }

search/VertexVisitor.hh

@@ -32,19 +32,20 @@ public:
   virtual VertexVisitor *copy() = 0;
   virtual void visit(Vertex *vertex) = 0;
   void operator()(Vertex *vertex) { visit(vertex); }
+  virtual void levelFinished() {}
 
 private:
   DISALLOW_COPY_AND_ASSIGN(VertexVisitor);
 };
 
-// Collector visited pins into a PinSet.
+// Collect visited pins into a PinSet.
 class VertexPinCollector : public VertexVisitor
 {
 public:
   VertexPinCollector(PinSet *pins);
   PinSet *pins() const { return pins_; }
   void visit(Vertex *vertex);
-  VertexVisitor *copy();
+  virtual VertexVisitor *copy();
 
 protected:
   PinSet *pins_;