| OLD | NEW |
| 1 //===- subzero/src/IceTimerTree.cpp - Pass timer defs ---------------------===// | 1 //===- subzero/src/IceTimerTree.cpp - Pass timer defs ---------------------===// |
| 2 // | 2 // |
| 3 // The Subzero Code Generator | 3 // The Subzero Code Generator |
| 4 // | 4 // |
| 5 // This file is distributed under the University of Illinois Open Source | 5 // This file is distributed under the University of Illinois Open Source |
| 6 // License. See LICENSE.TXT for details. | 6 // License. See LICENSE.TXT for details. |
| 7 // | 7 // |
| 8 //===----------------------------------------------------------------------===// | 8 //===----------------------------------------------------------------------===// |
| 9 /// | 9 /// |
| 10 /// \file | 10 /// \file |
| 11 /// This file defines the TimerTree class, which tracks flat and | 11 /// This file defines the TimerTree class, which tracks flat and cumulative |
| 12 /// cumulative execution time collection of call chains. | 12 /// execution time collection of call chains. |
| 13 /// | 13 /// |
| 14 //===----------------------------------------------------------------------===// | 14 //===----------------------------------------------------------------------===// |
| 15 | 15 |
| 16 #include "IceTimerTree.h" | 16 #include "IceTimerTree.h" |
| 17 | 17 |
| 18 #include "IceDefs.h" | 18 #include "IceDefs.h" |
| 19 | 19 |
| 20 #pragma clang diagnostic push | 20 #pragma clang diagnostic push |
| 21 #pragma clang diagnostic ignored "-Wunused-parameter" | 21 #pragma clang diagnostic ignored "-Wunused-parameter" |
| 22 #include "llvm/Support/Timer.h" | 22 #include "llvm/Support/Timer.h" |
| (...skipping 11 matching lines...) Expand all Loading... |
| 34 LeafCounts.resize(TT__num); | 34 LeafCounts.resize(TT__num); |
| 35 #define STR(s) #s | 35 #define STR(s) #s |
| 36 #define X(tag) \ | 36 #define X(tag) \ |
| 37 IDs[TT_##tag] = STR(tag); \ | 37 IDs[TT_##tag] = STR(tag); \ |
| 38 IDsIndex[STR(tag)] = TT_##tag; | 38 IDsIndex[STR(tag)] = TT_##tag; |
| 39 TIMERTREE_TABLE; | 39 TIMERTREE_TABLE; |
| 40 #undef X | 40 #undef X |
| 41 #undef STR | 41 #undef STR |
| 42 } | 42 } |
| 43 | 43 |
| 44 // Returns the unique timer ID for the given Name, creating a new ID | 44 // Returns the unique timer ID for the given Name, creating a new ID if needed. |
| 45 // if needed. | |
| 46 TimerIdT TimerStack::getTimerID(const IceString &Name) { | 45 TimerIdT TimerStack::getTimerID(const IceString &Name) { |
| 47 if (!BuildDefs::dump()) | 46 if (!BuildDefs::dump()) |
| 48 return 0; | 47 return 0; |
| 49 if (IDsIndex.find(Name) == IDsIndex.end()) { | 48 if (IDsIndex.find(Name) == IDsIndex.end()) { |
| 50 IDsIndex[Name] = IDs.size(); | 49 IDsIndex[Name] = IDs.size(); |
| 51 IDs.push_back(Name); | 50 IDs.push_back(Name); |
| 52 LeafTimes.push_back(decltype(LeafTimes)::value_type()); | 51 LeafTimes.push_back(decltype(LeafTimes)::value_type()); |
| 53 LeafCounts.push_back(decltype(LeafCounts)::value_type()); | 52 LeafCounts.push_back(decltype(LeafCounts)::value_type()); |
| 54 } | 53 } |
| 55 return IDsIndex[Name]; | 54 return IDsIndex[Name]; |
| 56 } | 55 } |
| 57 | 56 |
| 58 // Creates a mapping from TimerIdT (leaf) values in the Src timer | 57 // Creates a mapping from TimerIdT (leaf) values in the Src timer stack into |
| 59 // stack into TimerIdT values in this timer stack. Creates new | 58 // TimerIdT values in this timer stack. Creates new entries in this timer stack |
| 60 // entries in this timer stack as needed. | 59 // as needed. |
| 61 TimerStack::TranslationType | 60 TimerStack::TranslationType |
| 62 TimerStack::translateIDsFrom(const TimerStack &Src) { | 61 TimerStack::translateIDsFrom(const TimerStack &Src) { |
| 63 size_t Size = Src.IDs.size(); | 62 size_t Size = Src.IDs.size(); |
| 64 TranslationType Mapping(Size); | 63 TranslationType Mapping(Size); |
| 65 for (TimerIdT i = 0; i < Size; ++i) { | 64 for (TimerIdT i = 0; i < Size; ++i) { |
| 66 Mapping[i] = getTimerID(Src.IDs[i]); | 65 Mapping[i] = getTimerID(Src.IDs[i]); |
| 67 } | 66 } |
| 68 return Mapping; | 67 return Mapping; |
| 69 } | 68 } |
| 70 | 69 |
| 71 // Merges two timer stacks, by combining and summing corresponding | 70 // Merges two timer stacks, by combining and summing corresponding entries. |
| 72 // entries. This timer stack is updated from Src. | 71 // This timer stack is updated from Src. |
| 73 void TimerStack::mergeFrom(const TimerStack &Src) { | 72 void TimerStack::mergeFrom(const TimerStack &Src) { |
| 74 if (!BuildDefs::dump()) | 73 if (!BuildDefs::dump()) |
| 75 return; | 74 return; |
| 76 TranslationType Mapping = translateIDsFrom(Src); | 75 TranslationType Mapping = translateIDsFrom(Src); |
| 77 TTindex SrcIndex = 0; | 76 TTindex SrcIndex = 0; |
| 78 for (const TimerTreeNode &SrcNode : Src.Nodes) { | 77 for (const TimerTreeNode &SrcNode : Src.Nodes) { |
| 79 // The first node is reserved as a sentinel, so avoid it. | 78 // The first node is reserved as a sentinel, so avoid it. |
| 80 if (SrcIndex > 0) { | 79 if (SrcIndex > 0) { |
| 81 // Find the full path to the Src node, translated to path | 80 // Find the full path to the Src node, translated to path components |
| 82 // components corresponding to this timer stack. | 81 // corresponding to this timer stack. |
| 83 PathType MyPath = Src.getPath(SrcIndex, Mapping); | 82 PathType MyPath = Src.getPath(SrcIndex, Mapping); |
| 84 // Find a node in this timer stack corresponding to the given | 83 // Find a node in this timer stack corresponding to the given path, |
| 85 // path, creating new interior nodes as necessary. | 84 // creating new interior nodes as necessary. |
| 86 TTindex MyIndex = findPath(MyPath); | 85 TTindex MyIndex = findPath(MyPath); |
| 87 Nodes[MyIndex].Time += SrcNode.Time; | 86 Nodes[MyIndex].Time += SrcNode.Time; |
| 88 Nodes[MyIndex].UpdateCount += SrcNode.UpdateCount; | 87 Nodes[MyIndex].UpdateCount += SrcNode.UpdateCount; |
| 89 } | 88 } |
| 90 ++SrcIndex; | 89 ++SrcIndex; |
| 91 } | 90 } |
| 92 for (TimerIdT i = 0; i < Src.LeafTimes.size(); ++i) { | 91 for (TimerIdT i = 0; i < Src.LeafTimes.size(); ++i) { |
| 93 LeafTimes[Mapping[i]] += Src.LeafTimes[i]; | 92 LeafTimes[Mapping[i]] += Src.LeafTimes[i]; |
| 94 LeafCounts[Mapping[i]] += Src.LeafCounts[i]; | 93 LeafCounts[Mapping[i]] += Src.LeafCounts[i]; |
| 95 } | 94 } |
| 96 StateChangeCount += Src.StateChangeCount; | 95 StateChangeCount += Src.StateChangeCount; |
| 97 } | 96 } |
| 98 | 97 |
| 99 // Constructs a path consisting of the sequence of leaf values leading | 98 // Constructs a path consisting of the sequence of leaf values leading to a |
| 100 // to a given node, with the Mapping translation applied to the leaf | 99 // given node, with the Mapping translation applied to the leaf values. The |
| 101 // values. The path ends up being in "reverse" order, i.e. from leaf | 100 // path ends up being in "reverse" order, i.e. from leaf to root. |
| 102 // to root. | |
| 103 TimerStack::PathType TimerStack::getPath(TTindex Index, | 101 TimerStack::PathType TimerStack::getPath(TTindex Index, |
| 104 const TranslationType &Mapping) const { | 102 const TranslationType &Mapping) const { |
| 105 PathType Path; | 103 PathType Path; |
| 106 while (Index) { | 104 while (Index) { |
| 107 Path.push_back(Mapping[Nodes[Index].Interior]); | 105 Path.push_back(Mapping[Nodes[Index].Interior]); |
| 108 assert(Nodes[Index].Parent < Index); | 106 assert(Nodes[Index].Parent < Index); |
| 109 Index = Nodes[Index].Parent; | 107 Index = Nodes[Index].Parent; |
| 110 } | 108 } |
| 111 return Path; | 109 return Path; |
| 112 } | 110 } |
| 113 | 111 |
| 114 // Given a parent node and a leaf ID, returns the index of the | 112 // Given a parent node and a leaf ID, returns the index of the parent's child |
| 115 // parent's child ID, creating a new node for the child as necessary. | 113 // ID, creating a new node for the child as necessary. |
| 116 TimerStack::TTindex TimerStack::getChildIndex(TimerStack::TTindex Parent, | 114 TimerStack::TTindex TimerStack::getChildIndex(TimerStack::TTindex Parent, |
| 117 TimerIdT ID) { | 115 TimerIdT ID) { |
| 118 if (Nodes[Parent].Children.size() <= ID) | 116 if (Nodes[Parent].Children.size() <= ID) |
| 119 Nodes[Parent].Children.resize(ID + 1); | 117 Nodes[Parent].Children.resize(ID + 1); |
| 120 if (Nodes[Parent].Children[ID] == 0) { | 118 if (Nodes[Parent].Children[ID] == 0) { |
| 121 TTindex Size = Nodes.size(); | 119 TTindex Size = Nodes.size(); |
| 122 Nodes[Parent].Children[ID] = Size; | 120 Nodes[Parent].Children[ID] = Size; |
| 123 Nodes.resize(Size + 1); | 121 Nodes.resize(Size + 1); |
| 124 Nodes[Size].Parent = Parent; | 122 Nodes[Size].Parent = Parent; |
| 125 Nodes[Size].Interior = ID; | 123 Nodes[Size].Interior = ID; |
| 126 } | 124 } |
| 127 return Nodes[Parent].Children[ID]; | 125 return Nodes[Parent].Children[ID]; |
| 128 } | 126 } |
| 129 | 127 |
| 130 // Finds a node in the timer stack corresponding to the given path, | 128 // Finds a node in the timer stack corresponding to the given path, creating |
| 131 // creating new interior nodes as necessary. | 129 // new interior nodes as necessary. |
| 132 TimerStack::TTindex TimerStack::findPath(const PathType &Path) { | 130 TimerStack::TTindex TimerStack::findPath(const PathType &Path) { |
| 133 TTindex CurIndex = 0; | 131 TTindex CurIndex = 0; |
| 134 // The path is in reverse order (leaf to root), so it needs to be | 132 // The path is in reverse order (leaf to root), so it needs to be followed in |
| 135 // followed in reverse. | 133 // reverse. |
| 136 for (TTindex Index : reverse_range(Path)) { | 134 for (TTindex Index : reverse_range(Path)) { |
| 137 CurIndex = getChildIndex(CurIndex, Index); | 135 CurIndex = getChildIndex(CurIndex, Index); |
| 138 } | 136 } |
| 139 assert(CurIndex); // shouldn't be the sentinel node | 137 assert(CurIndex); // shouldn't be the sentinel node |
| 140 return CurIndex; | 138 return CurIndex; |
| 141 } | 139 } |
| 142 | 140 |
| 143 // Pushes a new marker onto the timer stack. | 141 // Pushes a new marker onto the timer stack. |
| 144 void TimerStack::push(TimerIdT ID) { | 142 void TimerStack::push(TimerIdT ID) { |
| 145 if (!BuildDefs::dump()) | 143 if (!BuildDefs::dump()) |
| 146 return; | 144 return; |
| 147 const bool UpdateCounts = false; | 145 const bool UpdateCounts = false; |
| 148 update(UpdateCounts); | 146 update(UpdateCounts); |
| 149 StackTop = getChildIndex(StackTop, ID); | 147 StackTop = getChildIndex(StackTop, ID); |
| 150 assert(StackTop); | 148 assert(StackTop); |
| 151 } | 149 } |
| 152 | 150 |
| 153 // Pops the top marker from the timer stack. Validates via assert() | 151 // Pops the top marker from the timer stack. Validates via assert() that the |
| 154 // that the expected marker is popped. | 152 // expected marker is popped. |
| 155 void TimerStack::pop(TimerIdT ID) { | 153 void TimerStack::pop(TimerIdT ID) { |
| 156 if (!BuildDefs::dump()) | 154 if (!BuildDefs::dump()) |
| 157 return; | 155 return; |
| 158 const bool UpdateCounts = true; | 156 const bool UpdateCounts = true; |
| 159 update(UpdateCounts); | 157 update(UpdateCounts); |
| 160 assert(StackTop); | 158 assert(StackTop); |
| 161 assert(Nodes[StackTop].Parent < StackTop); | 159 assert(Nodes[StackTop].Parent < StackTop); |
| 162 // Verify that the expected ID is being popped. | 160 // Verify that the expected ID is being popped. |
| 163 assert(Nodes[StackTop].Interior == ID); | 161 assert(Nodes[StackTop].Interior == ID); |
| 164 (void)ID; | 162 (void)ID; |
| 165 // Verify that the parent's child points to the current stack top. | 163 // Verify that the parent's child points to the current stack top. |
| 166 assert(Nodes[Nodes[StackTop].Parent].Children[ID] == StackTop); | 164 assert(Nodes[Nodes[StackTop].Parent].Children[ID] == StackTop); |
| 167 StackTop = Nodes[StackTop].Parent; | 165 StackTop = Nodes[StackTop].Parent; |
| 168 } | 166 } |
| 169 | 167 |
| 170 // At a state change (e.g. push or pop), updates the flat and | 168 // At a state change (e.g. push or pop), updates the flat and cumulative |
| 171 // cumulative timings for everything on the timer stack. | 169 // timings for everything on the timer stack. |
| 172 void TimerStack::update(bool UpdateCounts) { | 170 void TimerStack::update(bool UpdateCounts) { |
| 173 if (!BuildDefs::dump()) | 171 if (!BuildDefs::dump()) |
| 174 return; | 172 return; |
| 175 ++StateChangeCount; | 173 ++StateChangeCount; |
| 176 // Whenever the stack is about to change, we grab the time delta | 174 // Whenever the stack is about to change, we grab the time delta since the |
| 177 // since the last change and add it to all active cumulative | 175 // last change and add it to all active cumulative elements and to the flat |
| 178 // elements and to the flat element for the top of the stack. | 176 // element for the top of the stack. |
| 179 double Current = timestamp(); | 177 double Current = timestamp(); |
| 180 double Delta = Current - LastTimestamp; | 178 double Delta = Current - LastTimestamp; |
| 181 if (StackTop) { | 179 if (StackTop) { |
| 182 TimerIdT Leaf = Nodes[StackTop].Interior; | 180 TimerIdT Leaf = Nodes[StackTop].Interior; |
| 183 if (Leaf >= LeafTimes.size()) { | 181 if (Leaf >= LeafTimes.size()) { |
| 184 LeafTimes.resize(Leaf + 1); | 182 LeafTimes.resize(Leaf + 1); |
| 185 LeafCounts.resize(Leaf + 1); | 183 LeafCounts.resize(Leaf + 1); |
| 186 } | 184 } |
| 187 LeafTimes[Leaf] += Delta; | 185 LeafTimes[Leaf] += Delta; |
| 188 if (UpdateCounts) | 186 if (UpdateCounts) |
| 189 ++LeafCounts[Leaf]; | 187 ++LeafCounts[Leaf]; |
| 190 } | 188 } |
| 191 TTindex Prefix = StackTop; | 189 TTindex Prefix = StackTop; |
| 192 while (Prefix) { | 190 while (Prefix) { |
| 193 Nodes[Prefix].Time += Delta; | 191 Nodes[Prefix].Time += Delta; |
| 194 // Only update a leaf node count, not the internal node counts. | 192 // Only update a leaf node count, not the internal node counts. |
| 195 if (UpdateCounts && Prefix == StackTop) | 193 if (UpdateCounts && Prefix == StackTop) |
| 196 ++Nodes[Prefix].UpdateCount; | 194 ++Nodes[Prefix].UpdateCount; |
| 197 TTindex Next = Nodes[Prefix].Parent; | 195 TTindex Next = Nodes[Prefix].Parent; |
| 198 assert(Next < Prefix); | 196 assert(Next < Prefix); |
| 199 Prefix = Next; | 197 Prefix = Next; |
| 200 } | 198 } |
| 201 // Capture the next timestamp *after* the updates are finished. | 199 // Capture the next timestamp *after* the updates are finished. This |
| 202 // This minimizes how much the timer can perturb the reported | 200 // minimizes how much the timer can perturb the reported timing. The numbers |
| 203 // timing. The numbers may not sum to 100%, and the missing amount | 201 // may not sum to 100%, and the missing amount is indicative of the overhead |
| 204 // is indicative of the overhead of timing. | 202 // of timing. |
| 205 LastTimestamp = timestamp(); | 203 LastTimestamp = timestamp(); |
| 206 } | 204 } |
| 207 | 205 |
| 208 void TimerStack::reset() { | 206 void TimerStack::reset() { |
| 209 if (!BuildDefs::dump()) | 207 if (!BuildDefs::dump()) |
| 210 return; | 208 return; |
| 211 StateChangeCount = 0; | 209 StateChangeCount = 0; |
| 212 FirstTimestamp = LastTimestamp = timestamp(); | 210 FirstTimestamp = LastTimestamp = timestamp(); |
| 213 LeafTimes.assign(LeafTimes.size(), 0); | 211 LeafTimes.assign(LeafTimes.size(), 0); |
| 214 LeafCounts.assign(LeafCounts.size(), 0); | 212 LeafCounts.assign(LeafCounts.size(), 0); |
| (...skipping 12 matching lines...) Expand all Loading... |
| 227 if (!BuildDefs::dump()) | 225 if (!BuildDefs::dump()) |
| 228 return; | 226 return; |
| 229 for (auto &I : reverse_range(Map)) { | 227 for (auto &I : reverse_range(Map)) { |
| 230 char buf[80]; | 228 char buf[80]; |
| 231 snprintf(buf, llvm::array_lengthof(buf), " %10.6f (%4.1f%%): ", I.first, | 229 snprintf(buf, llvm::array_lengthof(buf), " %10.6f (%4.1f%%): ", I.first, |
| 232 I.first * 100 / TotalTime); | 230 I.first * 100 / TotalTime); |
| 233 Str << buf << I.second << "\n"; | 231 Str << buf << I.second << "\n"; |
| 234 } | 232 } |
| 235 } | 233 } |
| 236 | 234 |
| 237 // Write a printf() format string into Buf[], in the format "[%5lu] ", | 235 // Write a printf() format string into Buf[], in the format "[%5lu] ", where |
| 238 // where "5" is actually the number of digits in MaxVal. E.g., | 236 // "5" is actually the number of digits in MaxVal. E.g., |
| 239 // MaxVal=0 ==> "[%1lu] " | 237 // MaxVal=0 ==> "[%1lu] " |
| 240 // MaxVal=5 ==> "[%1lu] " | 238 // MaxVal=5 ==> "[%1lu] " |
| 241 // MaxVal=9876 ==> "[%4lu] " | 239 // MaxVal=9876 ==> "[%4lu] " |
| 242 void makePrintfFormatString(char *Buf, size_t BufLen, size_t MaxVal) { | 240 void makePrintfFormatString(char *Buf, size_t BufLen, size_t MaxVal) { |
| 243 if (!BuildDefs::dump()) | 241 if (!BuildDefs::dump()) |
| 244 return; | 242 return; |
| 245 int NumDigits = 0; | 243 int NumDigits = 0; |
| 246 do { | 244 do { |
| 247 ++NumDigits; | 245 ++NumDigits; |
| 248 MaxVal /= 10; | 246 MaxVal /= 10; |
| (...skipping 54 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 303 dumpHelper(Str, FlatMap, TotalTime); | 301 dumpHelper(Str, FlatMap, TotalTime); |
| 304 Str << "Number of timer updates: " << StateChangeCount << "\n"; | 302 Str << "Number of timer updates: " << StateChangeCount << "\n"; |
| 305 } | 303 } |
| 306 | 304 |
| 307 double TimerStack::timestamp() { | 305 double TimerStack::timestamp() { |
| 308 // TODO: Implement in terms of std::chrono for C++11. | 306 // TODO: Implement in terms of std::chrono for C++11. |
| 309 return llvm::TimeRecord::getCurrentTime(false).getWallTime(); | 307 return llvm::TimeRecord::getCurrentTime(false).getWallTime(); |
| 310 } | 308 } |
| 311 | 309 |
| 312 } // end of namespace Ice | 310 } // end of namespace Ice |
| OLD | NEW |