Merge pull request #261 from eder-matheus/master

CMakeLists.txt

@@ -38,7 +38,7 @@ project(STA VERSION 2.6.0
 
 option(CUDD_DIR "CUDD BDD package directory")
 option(USE_TCL_READLINE "Use TCL readline package" ON)
-option(USE_SANITIZE "Compile with santize address enabled")
+option(ENABLE_TSAN "Compile with thread santizer enabled" OFF)
 
 # Turn on to debug compiler args.
 set(CMAKE_VERBOSE_MAKEFILE OFF)
@@ -550,8 +550,8 @@ endif()
 # common to gcc/clang
 set(CXX_FLAGS -Wall -Wextra -pedantic -Wcast-qual -Wredundant-decls -Wformat-security)
 
-if(USE_SANITIZE)
-  message(STATUS "Sanitize: ${USE_SANITIZE}")
+if(ENABLE_TSAN)
+  message(STATUS "Thread sanitizer: ${ENABLE_TSAN}")
   set(CXX_FLAGS "${CXX_FLAGS};-fsanitize=thread")
   set(CMAKE_EXE_LINKER_FLAGS "-fsanitize=thread")
 endif()

dcalc/GraphDelayCalc.cc

@@ -685,13 +685,15 @@ GraphDelayCalc::findDriverDelays(Vertex *drvr_vertex,
                                  LoadPinIndexMap &load_pin_index_map)
 {
   MultiDrvrNet *multi_drvr = findMultiDrvrNet(drvr_vertex);
-  if (multi_drvr == nullptr
-      || (multi_drvr
-          && (!multi_drvr->parallelGates(network_)
-              || drvr_vertex == multi_drvr->dcalcDrvr()))) {
+  if (multi_drvr == nullptr) {
     initLoadSlews(drvr_vertex);
     findDriverDelays1(drvr_vertex, multi_drvr, arc_delay_calc, load_pin_index_map);
   }
+  else if (drvr_vertex == multi_drvr->dcalcDrvr()) {
+    initLoadSlews(drvr_vertex);
+    for (Vertex *drvr : multi_drvr->drvrs())
+      findDriverDelays1(drvr, multi_drvr, arc_delay_calc, load_pin_index_map);
+  }
   arc_delay_calc_->finishDrvrPin();
 }
 
@@ -825,16 +827,7 @@ GraphDelayCalc::findDriverDelays1(Vertex *drvr_vertex,
                                   LoadPinIndexMap &load_pin_index_map)
 {
   initSlew(drvr_vertex);
-  if (multi_drvr
-      && multi_drvr->parallelGates(network_)) {
-    // Only init on the trigger driver.
-    if (drvr_vertex == multi_drvr->dcalcDrvr()) {
-      for (auto vertex : multi_drvr->drvrs())
-        initWireDelays(vertex);
-    }
-  }
-  else
-    initWireDelays(drvr_vertex);
+  initWireDelays(drvr_vertex);
   bool delay_changed = false;
   array<bool, RiseFall::index_count> delay_exists = {false, false};
   VertexInEdgeIterator edge_iter(drvr_vertex, graph_);

doc/CodingGuidelines.txt

@@ -61,12 +61,20 @@ instead, write this:
        fatal ("virtual memory exhausted");
 
 
-Use braces around if/for bodies that are more than one line.
-IE, 
+Do not use braces around if/for that are one line.
+
   if (pred)
-    for (int i = 0; i < len; i++) {  // this body should be in {}'s
+    bar = 1;
+  else
+    bar = 3;
+
+Use braces around if/for bodies that are more than one line.
+
+  if (pred) {
+    for (int i = 0; i < len; i++) {
        ...
     }
+  }
 
 Add a default clause to all switches calling switchCaseNotHandled:
 

include/sta/ArrayTable.hh

@@ -26,6 +26,7 @@
 
 #include <cstring> // memcpy
 #include <vector>
+#include <atomic>
 
 #include "ObjectId.hh"
 #include "Error.hh"
@@ -77,8 +78,7 @@ private:
   // Don't use std::vector so growing blocks_ can be thread safe.
   size_t blocks_size_;
   size_t blocks_capacity_;
-  ArrayBlock<TYPE>* *blocks_;
-  ArrayBlock<TYPE>* *prev_blocks_;
+  std::atomic<ArrayBlock<TYPE>**> blocks_;
   // Linked list of free arrays indexed by array size.
   std::vector<ObjectId> free_list_;
   static constexpr ObjectId idx_mask_ = block_size - 1;
@@ -91,8 +91,7 @@ ArrayTable<TYPE>::ArrayTable() :
   free_idx_(object_idx_null),
   blocks_size_(0),
   blocks_capacity_(1024),
-  blocks_(new ArrayBlock<TYPE>*[blocks_capacity_]),
-  prev_blocks_(nullptr)
+  blocks_(new ArrayBlock<TYPE>*[blocks_capacity_])
 {
 }
 
@@ -101,7 +100,6 @@ ArrayTable<TYPE>::~ArrayTable()
 {
   deleteBlocks();
   delete [] blocks_;
-  delete [] prev_blocks_;
 }
 
 template <class TYPE>
@@ -173,10 +171,6 @@ ArrayTable<TYPE>::pushBlock(ArrayBlock<TYPE> *block)
     size_t new_capacity = blocks_capacity_ * 1.5;
     ArrayBlock<TYPE>** new_blocks = new ArrayBlock<TYPE>*[new_capacity];
     memcpy(new_blocks, blocks_, blocks_capacity_ * sizeof(ArrayBlock<TYPE>*));
-    if (prev_blocks_)
-      delete [] prev_blocks_;
-    // Preserve block array for other threads to reference.
-    prev_blocks_ = blocks_;
     blocks_ = new_blocks;
     blocks_capacity_ = new_capacity;
   }

include/sta/Graph.hh

@@ -25,6 +25,7 @@
 #pragma once
 
 #include <mutex>
+#include <atomic>
 
 #include "Iterator.hh"
 #include "Map.hh"
@@ -365,14 +366,12 @@ protected:
   EdgeId in_edges_;		// Edges to this vertex.
   EdgeId out_edges_;		// Edges from this vertex.
 
-  // 32 bits
+  // 28 bits
   unsigned int tag_group_index_:tag_group_index_bits; // 24
-  // Each bit corresponds to a different BFS queue.
-  unsigned int bfs_in_queue_:int(BfsIndex::bits); // 4
-  unsigned int slew_annotated_:slew_annotated_bits;
+  unsigned int slew_annotated_:slew_annotated_bits;  // 4
 
   // 32 bits
-  unsigned int level_:Graph::vertex_level_bits;
+  unsigned int level_:Graph::vertex_level_bits; // 24
   // Levelization search state.
   // LevelColor gcc barfs if this is dcl'd.
   unsigned color_:2;
@@ -383,6 +382,9 @@ protected:
   bool is_bidirect_drvr_:1;
   bool is_reg_clk_:1;
 
+  // Each bit corresponds to a different BFS queue.
+  std::atomic<uint8_t> bfs_in_queue_; // 4
+
   // 15 bits
   bool is_disabled_constraint_:1;
   bool is_gated_clk_enable_:1;
@@ -394,7 +396,7 @@ protected:
   bool has_downstream_clk_pin_:1;
   bool crpr_path_pruning_disabled_:1;
   bool requireds_pruned_:1;
-  unsigned object_idx_:VertexTable::idx_bits;
+  unsigned object_idx_:VertexTable::idx_bits; // 7
 
 private:
   friend class Graph;

include/sta/Liberty.hh

@@ -25,6 +25,7 @@
 #pragma once
 
 #include <mutex>
+#include <atomic>
 
 #include "MinMax.hh"
 #include "RiseFallMinMax.hh"
@@ -644,7 +645,7 @@ protected:
   bool leakage_power_exists_;
   LibertyPgPortMap pg_port_map_;
   bool has_internal_ports_;
-  bool have_voltage_waveforms_;
+  std::atomic<bool> have_voltage_waveforms_;
   std::mutex waveform_lock_;
   const char *footprint_;
   const char *user_function_class_;

include/sta/Search.hh

@@ -25,6 +25,7 @@
 #pragma once
 
 #include <mutex>
+#include <atomic>
 
 #include "MinMax.hh"
 #include "UnorderedSet.hh"
@@ -591,15 +592,13 @@ protected:
   TagSet *tag_set_;
   // Entries in tags_ may be missing where previous filter tags were deleted.
   TagIndex tag_capacity_;
-  Tag **tags_;
-  Tag **tags_prev_;
+  std::atomic<Tag **> tags_;
   TagIndex tag_next_;
   // Holes in tags_ left by deleting filter tags.
   std::vector<TagIndex> tag_free_indices_;
   std::mutex tag_lock_;
   TagGroupSet *tag_group_set_;
-  TagGroup **tag_groups_;
-  TagGroup **tag_groups_prev_;
+  std::atomic<TagGroup **> tag_groups_;
   TagGroupIndex tag_group_next_;
   // Holes in tag_groups_ left by deleting filter tag groups.
   std::vector<TagIndex> tag_group_free_indices_;

search/Search.cc

@@ -240,15 +240,13 @@ Search::init(StaState *sta)
   worst_slacks_ = nullptr;
   arrival_iter_ = new BfsFwdIterator(BfsIndex::arrival, nullptr, sta);
   required_iter_ = new BfsBkwdIterator(BfsIndex::required, search_adj_, sta);
-  tag_capacity_ = 127;
+  tag_capacity_ = 128;
   tag_set_ = new TagSet(tag_capacity_);
   clk_info_set_ = new ClkInfoSet(ClkInfoLess(sta));
   tag_next_ = 0;
   tags_ = new Tag*[tag_capacity_];
-  tags_prev_ = nullptr;
-  tag_group_capacity_ = 127;
+  tag_group_capacity_ = tag_capacity_;
   tag_groups_ = new TagGroup*[tag_group_capacity_];
-  tag_groups_prev_ = nullptr;
   tag_group_next_ = 0;
   tag_group_set_ = new TagGroupSet(tag_group_capacity_);
   pending_latch_outputs_ = new VertexSet(graph_);
@@ -280,9 +278,7 @@ Search::~Search()
   delete tag_set_;
   delete clk_info_set_;
   delete [] tags_;
-  delete [] tags_prev_;
   delete [] tag_groups_;
-  delete [] tag_groups_prev_;
   delete tag_group_set_;
   delete search_adj_;
   delete eval_pred_;
@@ -589,13 +585,14 @@ Search::deleteFilterTags()
 void
 Search::deleteFilterClkInfos()
 {
-  ClkInfoSet::Iterator clk_info_iter(clk_info_set_);
-  while (clk_info_iter.hasNext()) {
-    ClkInfo *clk_info = clk_info_iter.next();
+  for (auto itr = clk_info_set_->cbegin(); itr != clk_info_set_->cend(); ) {
+    ClkInfo *clk_info = *itr;
     if (clk_info->refsFilter(this)) {
-      clk_info_set_->erase(clk_info);
+      itr = clk_info_set_->erase(itr);
       delete clk_info;
     }
+    else
+    itr++;
   }
 }
 
@@ -2669,15 +2666,11 @@ Search::findTagGroup(TagGroupBldr *tag_bldr)
     // If tag_groups_ needs to grow make the new array and copy the
     // contents into it before updating tags_groups_ so that other threads
     // can use Search::tagGroup(TagGroupIndex) without returning gubbish.
-    // std::vector doesn't seem to follow this protocol so multi-thread
-    // search fails occasionally if a vector is used for tag_groups_.
     if (tag_group_next_ == tag_group_capacity_) {
       TagGroupIndex tag_capacity = tag_group_capacity_ * 2;
       TagGroup **tag_groups = new TagGroup*[tag_capacity];
       memcpy(tag_groups, tag_groups_,
              tag_group_capacity_ * sizeof(TagGroup*));
-      delete [] tag_groups_prev_;
-      tag_groups_prev_ = tag_groups_;
       tag_groups_ = tag_groups;
       tag_group_capacity_ = tag_capacity;
       tag_group_set_->reserve(tag_capacity);
@@ -2910,14 +2903,10 @@ Search::findTag(const RiseFall *rf,
     // If tags_ needs to grow make the new array and copy the
     // contents into it before updating tags_ so that other threads
     // can use Search::tag(TagIndex) without returning gubbish.
-    // std::vector doesn't seem to follow this protocol so multi-thread
-    // search fails occasionally if a vector is used for tags_.
     if (tag_next_ == tag_capacity_) {
       TagIndex tag_capacity = tag_capacity_ * 2;
       Tag **tags = new Tag*[tag_capacity];
       memcpy(tags, tags_, tag_capacity_ * sizeof(Tag*));
-      delete [] tags_prev_;
-      tags_prev_ = tags_;
       tags_ = tags;
       tag_capacity_ = tag_capacity;
       tag_set_->reserve(tag_capacity);