Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 1 | //===- CallGraphSort.cpp --------------------------------------------------===// |
| 2 | // |
| 3 | // The LLVM Linker |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | /// |
| 10 | /// Implementation of Call-Chain Clustering from: Optimizing Function Placement |
| 11 | /// for Large-Scale Data-Center Applications |
| 12 | /// https://research.fb.com/wp-content/uploads/2017/01/cgo2017-hfsort-final1.pdf |
| 13 | /// |
| 14 | /// The goal of this algorithm is to improve runtime performance of the final |
| 15 | /// executable by arranging code sections such that page table and i-cache |
| 16 | /// misses are minimized. |
| 17 | /// |
| 18 | /// Definitions: |
| 19 | /// * Cluster |
| 20 | /// * An ordered list of input sections which are layed out as a unit. At the |
| 21 | /// beginning of the algorithm each input section has its own cluster and |
| 22 | /// the weight of the cluster is the sum of the weight of all incomming |
| 23 | /// edges. |
| 24 | /// * Call-Chain Clustering (C³) Heuristic |
| 25 | /// * Defines when and how clusters are combined. Pick the highest weighted |
| 26 | /// input section then add it to its most likely predecessor if it wouldn't |
| 27 | /// penalize it too much. |
| 28 | /// * Density |
| 29 | /// * The weight of the cluster divided by the size of the cluster. This is a |
| 30 | /// proxy for the ammount of execution time spent per byte of the cluster. |
| 31 | /// |
| 32 | /// It does so given a call graph profile by the following: |
| 33 | /// * Build a weighted call graph from the call graph profile |
| 34 | /// * Sort input sections by weight |
| 35 | /// * For each input section starting with the highest weight |
| 36 | /// * Find its most likely predecessor cluster |
| 37 | /// * Check if the combined cluster would be too large, or would have too low |
| 38 | /// a density. |
| 39 | /// * If not, then combine the clusters. |
| 40 | /// * Sort non-empty clusters by density |
| 41 | /// |
| 42 | //===----------------------------------------------------------------------===// |
| 43 | |
| 44 | #include "CallGraphSort.h" |
| 45 | #include "OutputSections.h" |
| 46 | #include "SymbolTable.h" |
| 47 | #include "Symbols.h" |
| 48 | |
| 49 | using namespace llvm; |
| 50 | using namespace lld; |
| 51 | using namespace lld::elf; |
| 52 | |
| 53 | namespace { |
| 54 | struct Edge { |
| 55 | int From; |
| 56 | uint64_t Weight; |
| 57 | }; |
| 58 | |
| 59 | struct Cluster { |
| 60 | Cluster(int Sec, size_t S) { |
| 61 | Sections.push_back(Sec); |
| 62 | Size = S; |
| 63 | } |
| 64 | |
| 65 | double getDensity() const { |
| 66 | if (Size == 0) |
| 67 | return 0; |
| 68 | return double(Weight) / double(Size); |
| 69 | } |
| 70 | |
| 71 | std::vector<int> Sections; |
| 72 | size_t Size = 0; |
| 73 | uint64_t Weight = 0; |
| 74 | uint64_t InitialWeight = 0; |
George Rimar | f0eedbc | 2018-08-28 08:49:40 +0000 | [diff] [blame] | 75 | Edge BestPred = {-1, 0}; |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 76 | }; |
| 77 | |
| 78 | class CallGraphSort { |
| 79 | public: |
| 80 | CallGraphSort(); |
| 81 | |
| 82 | DenseMap<const InputSectionBase *, int> run(); |
| 83 | |
| 84 | private: |
| 85 | std::vector<Cluster> Clusters; |
| 86 | std::vector<const InputSectionBase *> Sections; |
| 87 | |
| 88 | void groupClusters(); |
| 89 | }; |
| 90 | |
| 91 | // Maximum ammount the combined cluster density can be worse than the original |
| 92 | // cluster to consider merging. |
| 93 | constexpr int MAX_DENSITY_DEGRADATION = 8; |
| 94 | |
| 95 | // Maximum cluster size in bytes. |
| 96 | constexpr uint64_t MAX_CLUSTER_SIZE = 1024 * 1024; |
| 97 | } // end anonymous namespace |
| 98 | |
| 99 | // Take the edge list in Config->CallGraphProfile, resolve symbol names to |
| 100 | // Symbols, and generate a graph between InputSections with the provided |
| 101 | // weights. |
| 102 | CallGraphSort::CallGraphSort() { |
| 103 | llvm::MapVector<std::pair<const InputSectionBase *, const InputSectionBase *>, |
| 104 | uint64_t> &Profile = Config->CallGraphProfile; |
| 105 | DenseMap<const InputSectionBase *, int> SecToCluster; |
| 106 | |
| 107 | auto GetOrCreateNode = [&](const InputSectionBase *IS) -> int { |
| 108 | auto Res = SecToCluster.insert(std::make_pair(IS, Clusters.size())); |
| 109 | if (Res.second) { |
| 110 | Sections.push_back(IS); |
| 111 | Clusters.emplace_back(Clusters.size(), IS->getSize()); |
| 112 | } |
| 113 | return Res.first->second; |
| 114 | }; |
| 115 | |
| 116 | // Create the graph. |
| 117 | for (const auto &C : Profile) { |
| 118 | const auto *FromSB = cast<InputSectionBase>(C.first.first->Repl); |
| 119 | const auto *ToSB = cast<InputSectionBase>(C.first.second->Repl); |
| 120 | uint64_t Weight = C.second; |
| 121 | |
| 122 | // Ignore edges between input sections belonging to different output |
| 123 | // sections. This is done because otherwise we would end up with clusters |
| 124 | // containing input sections that can't actually be placed adjacently in the |
| 125 | // output. This messes with the cluster size and density calculations. We |
| 126 | // would also end up moving input sections in other output sections without |
| 127 | // moving them closer to what calls them. |
| 128 | if (FromSB->getOutputSection() != ToSB->getOutputSection()) |
| 129 | continue; |
| 130 | |
| 131 | int From = GetOrCreateNode(FromSB); |
| 132 | int To = GetOrCreateNode(ToSB); |
| 133 | |
| 134 | Clusters[To].Weight += Weight; |
| 135 | |
| 136 | if (From == To) |
| 137 | continue; |
| 138 | |
George Rimar | f0eedbc | 2018-08-28 08:49:40 +0000 | [diff] [blame] | 139 | // Remember the best edge. |
| 140 | Cluster &ToC = Clusters[To]; |
| 141 | if (ToC.BestPred.From == -1 || ToC.BestPred.Weight < Weight) { |
| 142 | ToC.BestPred.From = From; |
| 143 | ToC.BestPred.Weight = Weight; |
| 144 | } |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 145 | } |
| 146 | for (Cluster &C : Clusters) |
| 147 | C.InitialWeight = C.Weight; |
| 148 | } |
| 149 | |
| 150 | // It's bad to merge clusters which would degrade the density too much. |
| 151 | static bool isNewDensityBad(Cluster &A, Cluster &B) { |
| 152 | double NewDensity = double(A.Weight + B.Weight) / double(A.Size + B.Size); |
| 153 | if (NewDensity < A.getDensity() / MAX_DENSITY_DEGRADATION) |
| 154 | return true; |
| 155 | return false; |
| 156 | } |
| 157 | |
| 158 | static void mergeClusters(Cluster &Into, Cluster &From) { |
| 159 | Into.Sections.insert(Into.Sections.end(), From.Sections.begin(), |
| 160 | From.Sections.end()); |
| 161 | Into.Size += From.Size; |
| 162 | Into.Weight += From.Weight; |
| 163 | From.Sections.clear(); |
| 164 | From.Size = 0; |
| 165 | From.Weight = 0; |
| 166 | } |
| 167 | |
| 168 | // Group InputSections into clusters using the Call-Chain Clustering heuristic |
| 169 | // then sort the clusters by density. |
| 170 | void CallGraphSort::groupClusters() { |
| 171 | std::vector<int> SortedSecs(Clusters.size()); |
| 172 | std::vector<Cluster *> SecToCluster(Clusters.size()); |
| 173 | |
George Rimar | a5cf8da | 2018-08-12 07:52:24 +0000 | [diff] [blame] | 174 | for (size_t I = 0; I < Clusters.size(); ++I) { |
| 175 | SortedSecs[I] = I; |
| 176 | SecToCluster[I] = &Clusters[I]; |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 177 | } |
| 178 | |
| 179 | std::stable_sort(SortedSecs.begin(), SortedSecs.end(), [&](int A, int B) { |
| 180 | return Clusters[B].getDensity() < Clusters[A].getDensity(); |
| 181 | }); |
| 182 | |
| 183 | for (int SI : SortedSecs) { |
| 184 | // Clusters[SI] is the same as SecToClusters[SI] here because it has not |
| 185 | // been merged into another cluster yet. |
| 186 | Cluster &C = Clusters[SI]; |
| 187 | |
George Rimar | f0eedbc | 2018-08-28 08:49:40 +0000 | [diff] [blame] | 188 | // Don't consider merging if the edge is unlikely. |
| 189 | if (C.BestPred.From == -1 || C.BestPred.Weight * 10 <= C.InitialWeight) |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 190 | continue; |
| 191 | |
George Rimar | f0eedbc | 2018-08-28 08:49:40 +0000 | [diff] [blame] | 192 | Cluster *PredC = SecToCluster[C.BestPred.From]; |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 193 | if (PredC == &C) |
| 194 | continue; |
| 195 | |
| 196 | if (C.Size + PredC->Size > MAX_CLUSTER_SIZE) |
| 197 | continue; |
| 198 | |
| 199 | if (isNewDensityBad(*PredC, C)) |
| 200 | continue; |
| 201 | |
| 202 | // NOTE: Consider using a disjoint-set to track section -> cluster mapping |
| 203 | // if this is ever slow. |
| 204 | for (int SI : C.Sections) |
| 205 | SecToCluster[SI] = PredC; |
| 206 | |
| 207 | mergeClusters(*PredC, C); |
| 208 | } |
| 209 | |
| 210 | // Remove empty or dead nodes. Invalidates all cluster indices. |
George Rimar | 97df22f | 2018-04-23 14:41:49 +0000 | [diff] [blame] | 211 | llvm::erase_if(Clusters, [](const Cluster &C) { |
| 212 | return C.Size == 0 || C.Sections.empty(); |
| 213 | }); |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 214 | |
| 215 | // Sort by density. |
George Rimar | de83cbf | 2018-04-24 09:55:39 +0000 | [diff] [blame] | 216 | std::stable_sort(Clusters.begin(), Clusters.end(), |
| 217 | [](const Cluster &A, const Cluster &B) { |
| 218 | return A.getDensity() > B.getDensity(); |
| 219 | }); |
Michael J. Spencer | b842725 | 2018-04-17 23:30:05 +0000 | [diff] [blame] | 220 | } |
| 221 | |
| 222 | DenseMap<const InputSectionBase *, int> CallGraphSort::run() { |
| 223 | groupClusters(); |
| 224 | |
| 225 | // Generate order. |
| 226 | llvm::DenseMap<const InputSectionBase *, int> OrderMap; |
| 227 | ssize_t CurOrder = 1; |
| 228 | |
| 229 | for (const Cluster &C : Clusters) |
| 230 | for (int SecIndex : C.Sections) |
| 231 | OrderMap[Sections[SecIndex]] = CurOrder++; |
| 232 | |
| 233 | return OrderMap; |
| 234 | } |
| 235 | |
| 236 | // Sort sections by the profile data provided by -callgraph-profile-file |
| 237 | // |
| 238 | // This first builds a call graph based on the profile data then merges sections |
| 239 | // according to the C³ huristic. All clusters are then sorted by a density |
| 240 | // metric to further improve locality. |
| 241 | DenseMap<const InputSectionBase *, int> elf::computeCallGraphProfileOrder() { |
| 242 | return CallGraphSort().run(); |
| 243 | } |