1 //===- CodeLayout.cpp - Implementation of code layout algorithms ----------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // The file implements "cache-aware" layout algorithms of basic blocks and 10 // functions in a binary. 11 // 12 // The algorithm tries to find a layout of nodes (basic blocks) of a given CFG 13 // optimizing jump locality and thus processor I-cache utilization. This is 14 // achieved via increasing the number of fall-through jumps and co-locating 15 // frequently executed nodes together. The name follows the underlying 16 // optimization problem, Extended-TSP, which is a generalization of classical 17 // (maximum) Traveling Salesmen Problem. 18 // 19 // The algorithm is a greedy heuristic that works with chains (ordered lists) 20 // of basic blocks. Initially all chains are isolated basic blocks. On every 21 // iteration, we pick a pair of chains whose merging yields the biggest increase 22 // in the ExtTSP score, which models how i-cache "friendly" a specific chain is. 23 // A pair of chains giving the maximum gain is merged into a new chain. The 24 // procedure stops when there is only one chain left, or when merging does not 25 // increase ExtTSP. In the latter case, the remaining chains are sorted by 26 // density in the decreasing order. 27 // 28 // An important aspect is the way two chains are merged. Unlike earlier 29 // algorithms (e.g., based on the approach of Pettis-Hansen), two 30 // chains, X and Y, are first split into three, X1, X2, and Y. Then we 31 // consider all possible ways of gluing the three chains (e.g., X1YX2, X1X2Y, 32 // X2X1Y, X2YX1, YX1X2, YX2X1) and choose the one producing the largest score. 33 // This improves the quality of the final result (the search space is larger) 34 // while keeping the implementation sufficiently fast. 35 // 36 // Reference: 37 // * A. Newell and S. Pupyrev, Improved Basic Block Reordering, 38 // IEEE Transactions on Computers, 2020 39 // https://arxiv.org/abs/1809.04676 40 // 41 //===----------------------------------------------------------------------===// 42 43 #include "llvm/Transforms/Utils/CodeLayout.h" 44 #include "llvm/Support/CommandLine.h" 45 #include "llvm/Support/Debug.h" 46 47 #include <cmath> 48 #include <set> 49 50 using namespace llvm; 51 using namespace llvm::codelayout; 52 53 #define DEBUG_TYPE "code-layout" 54 55 namespace llvm { 56 cl::opt<bool> EnableExtTspBlockPlacement( 57 "enable-ext-tsp-block-placement", cl::Hidden, cl::init(false), 58 cl::desc("Enable machine block placement based on the ext-tsp model, " 59 "optimizing I-cache utilization.")); 60 61 cl::opt<bool> ApplyExtTspWithoutProfile( 62 "ext-tsp-apply-without-profile", 63 cl::desc("Whether to apply ext-tsp placement for instances w/o profile"), 64 cl::init(true), cl::Hidden); 65 } // namespace llvm 66 67 // Algorithm-specific params for Ext-TSP. The values are tuned for the best 68 // performance of large-scale front-end bound binaries. 69 static cl::opt<double> ForwardWeightCond( 70 "ext-tsp-forward-weight-cond", cl::ReallyHidden, cl::init(0.1), 71 cl::desc("The weight of conditional forward jumps for ExtTSP value")); 72 73 static cl::opt<double> ForwardWeightUncond( 74 "ext-tsp-forward-weight-uncond", cl::ReallyHidden, cl::init(0.1), 75 cl::desc("The weight of unconditional forward jumps for ExtTSP value")); 76 77 static cl::opt<double> BackwardWeightCond( 78 "ext-tsp-backward-weight-cond", cl::ReallyHidden, cl::init(0.1), 79 cl::desc("The weight of conditional backward jumps for ExtTSP value")); 80 81 static cl::opt<double> BackwardWeightUncond( 82 "ext-tsp-backward-weight-uncond", cl::ReallyHidden, cl::init(0.1), 83 cl::desc("The weight of unconditional backward jumps for ExtTSP value")); 84 85 static cl::opt<double> FallthroughWeightCond( 86 "ext-tsp-fallthrough-weight-cond", cl::ReallyHidden, cl::init(1.0), 87 cl::desc("The weight of conditional fallthrough jumps for ExtTSP value")); 88 89 static cl::opt<double> FallthroughWeightUncond( 90 "ext-tsp-fallthrough-weight-uncond", cl::ReallyHidden, cl::init(1.05), 91 cl::desc("The weight of unconditional fallthrough jumps for ExtTSP value")); 92 93 static cl::opt<unsigned> ForwardDistance( 94 "ext-tsp-forward-distance", cl::ReallyHidden, cl::init(1024), 95 cl::desc("The maximum distance (in bytes) of a forward jump for ExtTSP")); 96 97 static cl::opt<unsigned> BackwardDistance( 98 "ext-tsp-backward-distance", cl::ReallyHidden, cl::init(640), 99 cl::desc("The maximum distance (in bytes) of a backward jump for ExtTSP")); 100 101 // The maximum size of a chain created by the algorithm. The size is bounded 102 // so that the algorithm can efficiently process extremely large instances. 103 static cl::opt<unsigned> 104 MaxChainSize("ext-tsp-max-chain-size", cl::ReallyHidden, cl::init(512), 105 cl::desc("The maximum size of a chain to create")); 106 107 // The maximum size of a chain for splitting. Larger values of the threshold 108 // may yield better quality at the cost of worsen run-time. 109 static cl::opt<unsigned> ChainSplitThreshold( 110 "ext-tsp-chain-split-threshold", cl::ReallyHidden, cl::init(128), 111 cl::desc("The maximum size of a chain to apply splitting")); 112 113 // The maximum ratio between densities of two chains for merging. 114 static cl::opt<double> MaxMergeDensityRatio( 115 "ext-tsp-max-merge-density-ratio", cl::ReallyHidden, cl::init(100), 116 cl::desc("The maximum ratio between densities of two chains for merging")); 117 118 // Algorithm-specific options for CDSort. 119 static cl::opt<unsigned> CacheEntries("cdsort-cache-entries", cl::ReallyHidden, 120 cl::desc("The size of the cache")); 121 122 static cl::opt<unsigned> CacheSize("cdsort-cache-size", cl::ReallyHidden, 123 cl::desc("The size of a line in the cache")); 124 125 static cl::opt<unsigned> 126 CDMaxChainSize("cdsort-max-chain-size", cl::ReallyHidden, 127 cl::desc("The maximum size of a chain to create")); 128 129 static cl::opt<double> DistancePower( 130 "cdsort-distance-power", cl::ReallyHidden, 131 cl::desc("The power exponent for the distance-based locality")); 132 133 static cl::opt<double> FrequencyScale( 134 "cdsort-frequency-scale", cl::ReallyHidden, 135 cl::desc("The scale factor for the frequency-based locality")); 136 137 namespace { 138 139 // Epsilon for comparison of doubles. 140 constexpr double EPS = 1e-8; 141 142 // Compute the Ext-TSP score for a given jump. 143 double jumpExtTSPScore(uint64_t JumpDist, uint64_t JumpMaxDist, uint64_t Count, 144 double Weight) { 145 if (JumpDist > JumpMaxDist) 146 return 0; 147 double Prob = 1.0 - static_cast<double>(JumpDist) / JumpMaxDist; 148 return Weight * Prob * Count; 149 } 150 151 // Compute the Ext-TSP score for a jump between a given pair of blocks, 152 // using their sizes, (estimated) addresses and the jump execution count. 153 double extTSPScore(uint64_t SrcAddr, uint64_t SrcSize, uint64_t DstAddr, 154 uint64_t Count, bool IsConditional) { 155 // Fallthrough 156 if (SrcAddr + SrcSize == DstAddr) { 157 return jumpExtTSPScore(0, 1, Count, 158 IsConditional ? FallthroughWeightCond 159 : FallthroughWeightUncond); 160 } 161 // Forward 162 if (SrcAddr + SrcSize < DstAddr) { 163 const uint64_t Dist = DstAddr - (SrcAddr + SrcSize); 164 return jumpExtTSPScore(Dist, ForwardDistance, Count, 165 IsConditional ? ForwardWeightCond 166 : ForwardWeightUncond); 167 } 168 // Backward 169 const uint64_t Dist = SrcAddr + SrcSize - DstAddr; 170 return jumpExtTSPScore(Dist, BackwardDistance, Count, 171 IsConditional ? BackwardWeightCond 172 : BackwardWeightUncond); 173 } 174 175 /// A type of merging two chains, X and Y. The former chain is split into 176 /// X1 and X2 and then concatenated with Y in the order specified by the type. 177 enum class MergeTypeT : int { X_Y, Y_X, X1_Y_X2, Y_X2_X1, X2_X1_Y }; 178 179 /// The gain of merging two chains, that is, the Ext-TSP score of the merge 180 /// together with the corresponding merge 'type' and 'offset'. 181 struct MergeGainT { 182 explicit MergeGainT() = default; 183 explicit MergeGainT(double Score, size_t MergeOffset, MergeTypeT MergeType) 184 : Score(Score), MergeOffset(MergeOffset), MergeType(MergeType) {} 185 186 double score() const { return Score; } 187 188 size_t mergeOffset() const { return MergeOffset; } 189 190 MergeTypeT mergeType() const { return MergeType; } 191 192 void setMergeType(MergeTypeT Ty) { MergeType = Ty; } 193 194 // Returns 'true' iff Other is preferred over this. 195 bool operator<(const MergeGainT &Other) const { 196 return (Other.Score > EPS && Other.Score > Score + EPS); 197 } 198 199 // Update the current gain if Other is preferred over this. 200 void updateIfLessThan(const MergeGainT &Other) { 201 if (*this < Other) 202 *this = Other; 203 } 204 205 private: 206 double Score{-1.0}; 207 size_t MergeOffset{0}; 208 MergeTypeT MergeType{MergeTypeT::X_Y}; 209 }; 210 211 struct JumpT; 212 struct ChainT; 213 struct ChainEdge; 214 215 /// A node in the graph, typically corresponding to a basic block in the CFG or 216 /// a function in the call graph. 217 struct NodeT { 218 NodeT(const NodeT &) = delete; 219 NodeT(NodeT &&) = default; 220 NodeT &operator=(const NodeT &) = delete; 221 NodeT &operator=(NodeT &&) = default; 222 223 explicit NodeT(size_t Index, uint64_t Size, uint64_t Count) 224 : Index(Index), Size(Size), ExecutionCount(Count) {} 225 226 bool isEntry() const { return Index == 0; } 227 228 // Check if Other is a successor of the node. 229 bool isSuccessor(const NodeT *Other) const; 230 231 // The total execution count of outgoing jumps. 232 uint64_t outCount() const; 233 234 // The total execution count of incoming jumps. 235 uint64_t inCount() const; 236 237 // The original index of the node in graph. 238 size_t Index{0}; 239 // The index of the node in the current chain. 240 size_t CurIndex{0}; 241 // The size of the node in the binary. 242 uint64_t Size{0}; 243 // The execution count of the node in the profile data. 244 uint64_t ExecutionCount{0}; 245 // The current chain of the node. 246 ChainT *CurChain{nullptr}; 247 // The offset of the node in the current chain. 248 mutable uint64_t EstimatedAddr{0}; 249 // Forced successor of the node in the graph. 250 NodeT *ForcedSucc{nullptr}; 251 // Forced predecessor of the node in the graph. 252 NodeT *ForcedPred{nullptr}; 253 // Outgoing jumps from the node. 254 std::vector<JumpT *> OutJumps; 255 // Incoming jumps to the node. 256 std::vector<JumpT *> InJumps; 257 }; 258 259 /// An arc in the graph, typically corresponding to a jump between two nodes. 260 struct JumpT { 261 JumpT(const JumpT &) = delete; 262 JumpT(JumpT &&) = default; 263 JumpT &operator=(const JumpT &) = delete; 264 JumpT &operator=(JumpT &&) = default; 265 266 explicit JumpT(NodeT *Source, NodeT *Target, uint64_t ExecutionCount) 267 : Source(Source), Target(Target), ExecutionCount(ExecutionCount) {} 268 269 // Source node of the jump. 270 NodeT *Source; 271 // Target node of the jump. 272 NodeT *Target; 273 // Execution count of the arc in the profile data. 274 uint64_t ExecutionCount{0}; 275 // Whether the jump corresponds to a conditional branch. 276 bool IsConditional{false}; 277 // The offset of the jump from the source node. 278 uint64_t Offset{0}; 279 }; 280 281 /// A chain (ordered sequence) of nodes in the graph. 282 struct ChainT { 283 ChainT(const ChainT &) = delete; 284 ChainT(ChainT &&) = default; 285 ChainT &operator=(const ChainT &) = delete; 286 ChainT &operator=(ChainT &&) = default; 287 288 explicit ChainT(uint64_t Id, NodeT *Node) 289 : Id(Id), ExecutionCount(Node->ExecutionCount), Size(Node->Size), 290 Nodes(1, Node) {} 291 292 size_t numBlocks() const { return Nodes.size(); } 293 294 double density() const { return ExecutionCount / Size; } 295 296 bool isEntry() const { return Nodes[0]->Index == 0; } 297 298 bool isCold() const { 299 for (NodeT *Node : Nodes) { 300 if (Node->ExecutionCount > 0) 301 return false; 302 } 303 return true; 304 } 305 306 ChainEdge *getEdge(ChainT *Other) const { 307 for (const auto &[Chain, ChainEdge] : Edges) { 308 if (Chain == Other) 309 return ChainEdge; 310 } 311 return nullptr; 312 } 313 314 void removeEdge(ChainT *Other) { 315 auto It = Edges.begin(); 316 while (It != Edges.end()) { 317 if (It->first == Other) { 318 Edges.erase(It); 319 return; 320 } 321 It++; 322 } 323 } 324 325 void addEdge(ChainT *Other, ChainEdge *Edge) { 326 Edges.push_back(std::make_pair(Other, Edge)); 327 } 328 329 void merge(ChainT *Other, std::vector<NodeT *> MergedBlocks) { 330 Nodes = std::move(MergedBlocks); 331 // Update the chain's data. 332 ExecutionCount += Other->ExecutionCount; 333 Size += Other->Size; 334 Id = Nodes[0]->Index; 335 // Update the node's data. 336 for (size_t Idx = 0; Idx < Nodes.size(); Idx++) { 337 Nodes[Idx]->CurChain = this; 338 Nodes[Idx]->CurIndex = Idx; 339 } 340 } 341 342 void mergeEdges(ChainT *Other); 343 344 void clear() { 345 Nodes.clear(); 346 Nodes.shrink_to_fit(); 347 Edges.clear(); 348 Edges.shrink_to_fit(); 349 } 350 351 // Unique chain identifier. 352 uint64_t Id; 353 // Cached ext-tsp score for the chain. 354 double Score{0}; 355 // The total execution count of the chain. Since the execution count of 356 // a basic block is uint64_t, using doubles here to avoid overflow. 357 double ExecutionCount{0}; 358 // The total size of the chain. 359 uint64_t Size{0}; 360 // Nodes of the chain. 361 std::vector<NodeT *> Nodes; 362 // Adjacent chains and corresponding edges (lists of jumps). 363 std::vector<std::pair<ChainT *, ChainEdge *>> Edges; 364 }; 365 366 /// An edge in the graph representing jumps between two chains. 367 /// When nodes are merged into chains, the edges are combined too so that 368 /// there is always at most one edge between a pair of chains. 369 struct ChainEdge { 370 ChainEdge(const ChainEdge &) = delete; 371 ChainEdge(ChainEdge &&) = default; 372 ChainEdge &operator=(const ChainEdge &) = delete; 373 ChainEdge &operator=(ChainEdge &&) = delete; 374 375 explicit ChainEdge(JumpT *Jump) 376 : SrcChain(Jump->Source->CurChain), DstChain(Jump->Target->CurChain), 377 Jumps(1, Jump) {} 378 379 ChainT *srcChain() const { return SrcChain; } 380 381 ChainT *dstChain() const { return DstChain; } 382 383 bool isSelfEdge() const { return SrcChain == DstChain; } 384 385 const std::vector<JumpT *> &jumps() const { return Jumps; } 386 387 void appendJump(JumpT *Jump) { Jumps.push_back(Jump); } 388 389 void moveJumps(ChainEdge *Other) { 390 Jumps.insert(Jumps.end(), Other->Jumps.begin(), Other->Jumps.end()); 391 Other->Jumps.clear(); 392 Other->Jumps.shrink_to_fit(); 393 } 394 395 void changeEndpoint(ChainT *From, ChainT *To) { 396 if (From == SrcChain) 397 SrcChain = To; 398 if (From == DstChain) 399 DstChain = To; 400 } 401 402 bool hasCachedMergeGain(ChainT *Src, ChainT *Dst) const { 403 return Src == SrcChain ? CacheValidForward : CacheValidBackward; 404 } 405 406 MergeGainT getCachedMergeGain(ChainT *Src, ChainT *Dst) const { 407 return Src == SrcChain ? CachedGainForward : CachedGainBackward; 408 } 409 410 void setCachedMergeGain(ChainT *Src, ChainT *Dst, MergeGainT MergeGain) { 411 if (Src == SrcChain) { 412 CachedGainForward = MergeGain; 413 CacheValidForward = true; 414 } else { 415 CachedGainBackward = MergeGain; 416 CacheValidBackward = true; 417 } 418 } 419 420 void invalidateCache() { 421 CacheValidForward = false; 422 CacheValidBackward = false; 423 } 424 425 void setMergeGain(MergeGainT Gain) { CachedGain = Gain; } 426 427 MergeGainT getMergeGain() const { return CachedGain; } 428 429 double gain() const { return CachedGain.score(); } 430 431 private: 432 // Source chain. 433 ChainT *SrcChain{nullptr}; 434 // Destination chain. 435 ChainT *DstChain{nullptr}; 436 // Original jumps in the binary with corresponding execution counts. 437 std::vector<JumpT *> Jumps; 438 // Cached gain value for merging the pair of chains. 439 MergeGainT CachedGain; 440 441 // Cached gain values for merging the pair of chains. Since the gain of 442 // merging (Src, Dst) and (Dst, Src) might be different, we store both values 443 // here and a flag indicating which of the options results in a higher gain. 444 // Cached gain values. 445 MergeGainT CachedGainForward; 446 MergeGainT CachedGainBackward; 447 // Whether the cached value must be recomputed. 448 bool CacheValidForward{false}; 449 bool CacheValidBackward{false}; 450 }; 451 452 bool NodeT::isSuccessor(const NodeT *Other) const { 453 for (JumpT *Jump : OutJumps) 454 if (Jump->Target == Other) 455 return true; 456 return false; 457 } 458 459 uint64_t NodeT::outCount() const { 460 uint64_t Count = 0; 461 for (JumpT *Jump : OutJumps) 462 Count += Jump->ExecutionCount; 463 return Count; 464 } 465 466 uint64_t NodeT::inCount() const { 467 uint64_t Count = 0; 468 for (JumpT *Jump : InJumps) 469 Count += Jump->ExecutionCount; 470 return Count; 471 } 472 473 void ChainT::mergeEdges(ChainT *Other) { 474 // Update edges adjacent to chain Other. 475 for (const auto &[DstChain, DstEdge] : Other->Edges) { 476 ChainT *TargetChain = DstChain == Other ? this : DstChain; 477 ChainEdge *CurEdge = getEdge(TargetChain); 478 if (CurEdge == nullptr) { 479 DstEdge->changeEndpoint(Other, this); 480 this->addEdge(TargetChain, DstEdge); 481 if (DstChain != this && DstChain != Other) 482 DstChain->addEdge(this, DstEdge); 483 } else { 484 CurEdge->moveJumps(DstEdge); 485 } 486 // Cleanup leftover edge. 487 if (DstChain != Other) 488 DstChain->removeEdge(Other); 489 } 490 } 491 492 using NodeIter = std::vector<NodeT *>::const_iterator; 493 static std::vector<NodeT *> EmptyList; 494 495 /// A wrapper around three concatenated vectors (chains) of nodes; it is used 496 /// to avoid extra instantiation of the vectors. 497 struct MergedNodesT { 498 MergedNodesT(NodeIter Begin1, NodeIter End1, 499 NodeIter Begin2 = EmptyList.begin(), 500 NodeIter End2 = EmptyList.end(), 501 NodeIter Begin3 = EmptyList.begin(), 502 NodeIter End3 = EmptyList.end()) 503 : Begin1(Begin1), End1(End1), Begin2(Begin2), End2(End2), Begin3(Begin3), 504 End3(End3) {} 505 506 template <typename F> void forEach(const F &Func) const { 507 for (auto It = Begin1; It != End1; It++) 508 Func(*It); 509 for (auto It = Begin2; It != End2; It++) 510 Func(*It); 511 for (auto It = Begin3; It != End3; It++) 512 Func(*It); 513 } 514 515 std::vector<NodeT *> getNodes() const { 516 std::vector<NodeT *> Result; 517 Result.reserve(std::distance(Begin1, End1) + std::distance(Begin2, End2) + 518 std::distance(Begin3, End3)); 519 Result.insert(Result.end(), Begin1, End1); 520 Result.insert(Result.end(), Begin2, End2); 521 Result.insert(Result.end(), Begin3, End3); 522 return Result; 523 } 524 525 const NodeT *getFirstNode() const { return *Begin1; } 526 527 private: 528 NodeIter Begin1; 529 NodeIter End1; 530 NodeIter Begin2; 531 NodeIter End2; 532 NodeIter Begin3; 533 NodeIter End3; 534 }; 535 536 /// A wrapper around two concatenated vectors (chains) of jumps. 537 struct MergedJumpsT { 538 MergedJumpsT(const std::vector<JumpT *> *Jumps1, 539 const std::vector<JumpT *> *Jumps2 = nullptr) { 540 assert(!Jumps1->empty() && "cannot merge empty jump list"); 541 JumpArray[0] = Jumps1; 542 JumpArray[1] = Jumps2; 543 } 544 545 template <typename F> void forEach(const F &Func) const { 546 for (auto Jumps : JumpArray) 547 if (Jumps != nullptr) 548 for (JumpT *Jump : *Jumps) 549 Func(Jump); 550 } 551 552 private: 553 std::array<const std::vector<JumpT *> *, 2> JumpArray{nullptr, nullptr}; 554 }; 555 556 /// Merge two chains of nodes respecting a given 'type' and 'offset'. 557 /// 558 /// If MergeType == 0, then the result is a concatenation of two chains. 559 /// Otherwise, the first chain is cut into two sub-chains at the offset, 560 /// and merged using all possible ways of concatenating three chains. 561 MergedNodesT mergeNodes(const std::vector<NodeT *> &X, 562 const std::vector<NodeT *> &Y, size_t MergeOffset, 563 MergeTypeT MergeType) { 564 // Split the first chain, X, into X1 and X2. 565 NodeIter BeginX1 = X.begin(); 566 NodeIter EndX1 = X.begin() + MergeOffset; 567 NodeIter BeginX2 = X.begin() + MergeOffset; 568 NodeIter EndX2 = X.end(); 569 NodeIter BeginY = Y.begin(); 570 NodeIter EndY = Y.end(); 571 572 // Construct a new chain from the three existing ones. 573 switch (MergeType) { 574 case MergeTypeT::X_Y: 575 return MergedNodesT(BeginX1, EndX2, BeginY, EndY); 576 case MergeTypeT::Y_X: 577 return MergedNodesT(BeginY, EndY, BeginX1, EndX2); 578 case MergeTypeT::X1_Y_X2: 579 return MergedNodesT(BeginX1, EndX1, BeginY, EndY, BeginX2, EndX2); 580 case MergeTypeT::Y_X2_X1: 581 return MergedNodesT(BeginY, EndY, BeginX2, EndX2, BeginX1, EndX1); 582 case MergeTypeT::X2_X1_Y: 583 return MergedNodesT(BeginX2, EndX2, BeginX1, EndX1, BeginY, EndY); 584 } 585 llvm_unreachable("unexpected chain merge type"); 586 } 587 588 /// The implementation of the ExtTSP algorithm. 589 class ExtTSPImpl { 590 public: 591 ExtTSPImpl(ArrayRef<uint64_t> NodeSizes, ArrayRef<uint64_t> NodeCounts, 592 ArrayRef<EdgeCount> EdgeCounts) 593 : NumNodes(NodeSizes.size()) { 594 initialize(NodeSizes, NodeCounts, EdgeCounts); 595 } 596 597 /// Run the algorithm and return an optimized ordering of nodes. 598 std::vector<uint64_t> run() { 599 // Pass 1: Merge nodes with their mutually forced successors 600 mergeForcedPairs(); 601 602 // Pass 2: Merge pairs of chains while improving the ExtTSP objective 603 mergeChainPairs(); 604 605 // Pass 3: Merge cold nodes to reduce code size 606 mergeColdChains(); 607 608 // Collect nodes from all chains 609 return concatChains(); 610 } 611 612 private: 613 /// Initialize the algorithm's data structures. 614 void initialize(const ArrayRef<uint64_t> &NodeSizes, 615 const ArrayRef<uint64_t> &NodeCounts, 616 const ArrayRef<EdgeCount> &EdgeCounts) { 617 // Initialize nodes. 618 AllNodes.reserve(NumNodes); 619 for (uint64_t Idx = 0; Idx < NumNodes; Idx++) { 620 uint64_t Size = std::max<uint64_t>(NodeSizes[Idx], 1ULL); 621 uint64_t ExecutionCount = NodeCounts[Idx]; 622 // The execution count of the entry node is set to at least one. 623 if (Idx == 0 && ExecutionCount == 0) 624 ExecutionCount = 1; 625 AllNodes.emplace_back(Idx, Size, ExecutionCount); 626 } 627 628 // Initialize jumps between the nodes. 629 SuccNodes.resize(NumNodes); 630 PredNodes.resize(NumNodes); 631 std::vector<uint64_t> OutDegree(NumNodes, 0); 632 AllJumps.reserve(EdgeCounts.size()); 633 for (auto Edge : EdgeCounts) { 634 ++OutDegree[Edge.src]; 635 // Ignore self-edges. 636 if (Edge.src == Edge.dst) 637 continue; 638 639 SuccNodes[Edge.src].push_back(Edge.dst); 640 PredNodes[Edge.dst].push_back(Edge.src); 641 if (Edge.count > 0) { 642 NodeT &PredNode = AllNodes[Edge.src]; 643 NodeT &SuccNode = AllNodes[Edge.dst]; 644 AllJumps.emplace_back(&PredNode, &SuccNode, Edge.count); 645 SuccNode.InJumps.push_back(&AllJumps.back()); 646 PredNode.OutJumps.push_back(&AllJumps.back()); 647 // Adjust execution counts. 648 PredNode.ExecutionCount = std::max(PredNode.ExecutionCount, Edge.count); 649 SuccNode.ExecutionCount = std::max(SuccNode.ExecutionCount, Edge.count); 650 } 651 } 652 for (JumpT &Jump : AllJumps) { 653 assert(OutDegree[Jump.Source->Index] > 0 && 654 "incorrectly computed out-degree of the block"); 655 Jump.IsConditional = OutDegree[Jump.Source->Index] > 1; 656 } 657 658 // Initialize chains. 659 AllChains.reserve(NumNodes); 660 HotChains.reserve(NumNodes); 661 for (NodeT &Node : AllNodes) { 662 // Create a chain. 663 AllChains.emplace_back(Node.Index, &Node); 664 Node.CurChain = &AllChains.back(); 665 if (Node.ExecutionCount > 0) 666 HotChains.push_back(&AllChains.back()); 667 } 668 669 // Initialize chain edges. 670 AllEdges.reserve(AllJumps.size()); 671 for (NodeT &PredNode : AllNodes) { 672 for (JumpT *Jump : PredNode.OutJumps) { 673 assert(Jump->ExecutionCount > 0 && "incorrectly initialized jump"); 674 NodeT *SuccNode = Jump->Target; 675 ChainEdge *CurEdge = PredNode.CurChain->getEdge(SuccNode->CurChain); 676 // This edge is already present in the graph. 677 if (CurEdge != nullptr) { 678 assert(SuccNode->CurChain->getEdge(PredNode.CurChain) != nullptr); 679 CurEdge->appendJump(Jump); 680 continue; 681 } 682 // This is a new edge. 683 AllEdges.emplace_back(Jump); 684 PredNode.CurChain->addEdge(SuccNode->CurChain, &AllEdges.back()); 685 SuccNode->CurChain->addEdge(PredNode.CurChain, &AllEdges.back()); 686 } 687 } 688 } 689 690 /// For a pair of nodes, A and B, node B is the forced successor of A, 691 /// if (i) all jumps (based on profile) from A goes to B and (ii) all jumps 692 /// to B are from A. Such nodes should be adjacent in the optimal ordering; 693 /// the method finds and merges such pairs of nodes. 694 void mergeForcedPairs() { 695 // Find forced pairs of blocks. 696 for (NodeT &Node : AllNodes) { 697 if (SuccNodes[Node.Index].size() == 1 && 698 PredNodes[SuccNodes[Node.Index][0]].size() == 1 && 699 SuccNodes[Node.Index][0] != 0) { 700 size_t SuccIndex = SuccNodes[Node.Index][0]; 701 Node.ForcedSucc = &AllNodes[SuccIndex]; 702 AllNodes[SuccIndex].ForcedPred = &Node; 703 } 704 } 705 706 // There might be 'cycles' in the forced dependencies, since profile 707 // data isn't 100% accurate. Typically this is observed in loops, when the 708 // loop edges are the hottest successors for the basic blocks of the loop. 709 // Break the cycles by choosing the node with the smallest index as the 710 // head. This helps to keep the original order of the loops, which likely 711 // have already been rotated in the optimized manner. 712 for (NodeT &Node : AllNodes) { 713 if (Node.ForcedSucc == nullptr || Node.ForcedPred == nullptr) 714 continue; 715 716 NodeT *SuccNode = Node.ForcedSucc; 717 while (SuccNode != nullptr && SuccNode != &Node) { 718 SuccNode = SuccNode->ForcedSucc; 719 } 720 if (SuccNode == nullptr) 721 continue; 722 // Break the cycle. 723 AllNodes[Node.ForcedPred->Index].ForcedSucc = nullptr; 724 Node.ForcedPred = nullptr; 725 } 726 727 // Merge nodes with their fallthrough successors. 728 for (NodeT &Node : AllNodes) { 729 if (Node.ForcedPred == nullptr && Node.ForcedSucc != nullptr) { 730 const NodeT *CurBlock = &Node; 731 while (CurBlock->ForcedSucc != nullptr) { 732 const NodeT *NextBlock = CurBlock->ForcedSucc; 733 mergeChains(Node.CurChain, NextBlock->CurChain, 0, MergeTypeT::X_Y); 734 CurBlock = NextBlock; 735 } 736 } 737 } 738 } 739 740 /// Merge pairs of chains while improving the ExtTSP objective. 741 void mergeChainPairs() { 742 /// Deterministically compare pairs of chains. 743 auto compareChainPairs = [](const ChainT *A1, const ChainT *B1, 744 const ChainT *A2, const ChainT *B2) { 745 return std::make_tuple(A1->Id, B1->Id) < std::make_tuple(A2->Id, B2->Id); 746 }; 747 748 while (HotChains.size() > 1) { 749 ChainT *BestChainPred = nullptr; 750 ChainT *BestChainSucc = nullptr; 751 MergeGainT BestGain; 752 // Iterate over all pairs of chains. 753 for (ChainT *ChainPred : HotChains) { 754 // Get candidates for merging with the current chain. 755 for (const auto &[ChainSucc, Edge] : ChainPred->Edges) { 756 // Ignore loop edges. 757 if (Edge->isSelfEdge()) 758 continue; 759 // Skip the merge if the combined chain violates the maximum specified 760 // size. 761 if (ChainPred->numBlocks() + ChainSucc->numBlocks() >= MaxChainSize) 762 continue; 763 // Don't merge the chains if they have vastly different densities. 764 // Skip the merge if the ratio between the densities exceeds 765 // MaxMergeDensityRatio. Smaller values of the option result in fewer 766 // merges, and hence, more chains. 767 const double ChainPredDensity = ChainPred->density(); 768 const double ChainSuccDensity = ChainSucc->density(); 769 assert(ChainPredDensity > 0.0 && ChainSuccDensity > 0.0 && 770 "incorrectly computed chain densities"); 771 auto [MinDensity, MaxDensity] = 772 std::minmax(ChainPredDensity, ChainSuccDensity); 773 const double Ratio = MaxDensity / MinDensity; 774 if (Ratio > MaxMergeDensityRatio) 775 continue; 776 777 // Compute the gain of merging the two chains. 778 MergeGainT CurGain = getBestMergeGain(ChainPred, ChainSucc, Edge); 779 if (CurGain.score() <= EPS) 780 continue; 781 782 if (BestGain < CurGain || 783 (std::abs(CurGain.score() - BestGain.score()) < EPS && 784 compareChainPairs(ChainPred, ChainSucc, BestChainPred, 785 BestChainSucc))) { 786 BestGain = CurGain; 787 BestChainPred = ChainPred; 788 BestChainSucc = ChainSucc; 789 } 790 } 791 } 792 793 // Stop merging when there is no improvement. 794 if (BestGain.score() <= EPS) 795 break; 796 797 // Merge the best pair of chains. 798 mergeChains(BestChainPred, BestChainSucc, BestGain.mergeOffset(), 799 BestGain.mergeType()); 800 } 801 } 802 803 /// Merge remaining nodes into chains w/o taking jump counts into 804 /// consideration. This allows to maintain the original node order in the 805 /// absence of profile data. 806 void mergeColdChains() { 807 for (size_t SrcBB = 0; SrcBB < NumNodes; SrcBB++) { 808 // Iterating in reverse order to make sure original fallthrough jumps are 809 // merged first; this might be beneficial for code size. 810 size_t NumSuccs = SuccNodes[SrcBB].size(); 811 for (size_t Idx = 0; Idx < NumSuccs; Idx++) { 812 size_t DstBB = SuccNodes[SrcBB][NumSuccs - Idx - 1]; 813 ChainT *SrcChain = AllNodes[SrcBB].CurChain; 814 ChainT *DstChain = AllNodes[DstBB].CurChain; 815 if (SrcChain != DstChain && !DstChain->isEntry() && 816 SrcChain->Nodes.back()->Index == SrcBB && 817 DstChain->Nodes.front()->Index == DstBB && 818 SrcChain->isCold() == DstChain->isCold()) { 819 mergeChains(SrcChain, DstChain, 0, MergeTypeT::X_Y); 820 } 821 } 822 } 823 } 824 825 /// Compute the Ext-TSP score for a given node order and a list of jumps. 826 double extTSPScore(const MergedNodesT &Nodes, 827 const MergedJumpsT &Jumps) const { 828 uint64_t CurAddr = 0; 829 Nodes.forEach([&](const NodeT *Node) { 830 Node->EstimatedAddr = CurAddr; 831 CurAddr += Node->Size; 832 }); 833 834 double Score = 0; 835 Jumps.forEach([&](const JumpT *Jump) { 836 const NodeT *SrcBlock = Jump->Source; 837 const NodeT *DstBlock = Jump->Target; 838 Score += ::extTSPScore(SrcBlock->EstimatedAddr, SrcBlock->Size, 839 DstBlock->EstimatedAddr, Jump->ExecutionCount, 840 Jump->IsConditional); 841 }); 842 return Score; 843 } 844 845 /// Compute the gain of merging two chains. 846 /// 847 /// The function considers all possible ways of merging two chains and 848 /// computes the one having the largest increase in ExtTSP objective. The 849 /// result is a pair with the first element being the gain and the second 850 /// element being the corresponding merging type. 851 MergeGainT getBestMergeGain(ChainT *ChainPred, ChainT *ChainSucc, 852 ChainEdge *Edge) const { 853 if (Edge->hasCachedMergeGain(ChainPred, ChainSucc)) 854 return Edge->getCachedMergeGain(ChainPred, ChainSucc); 855 856 assert(!Edge->jumps().empty() && "trying to merge chains w/o jumps"); 857 // Precompute jumps between ChainPred and ChainSucc. 858 ChainEdge *EdgePP = ChainPred->getEdge(ChainPred); 859 MergedJumpsT Jumps(&Edge->jumps(), EdgePP ? &EdgePP->jumps() : nullptr); 860 861 // This object holds the best chosen gain of merging two chains. 862 MergeGainT Gain = MergeGainT(); 863 864 /// Given a merge offset and a list of merge types, try to merge two chains 865 /// and update Gain with a better alternative. 866 auto tryChainMerging = [&](size_t Offset, 867 const std::vector<MergeTypeT> &MergeTypes) { 868 // Skip merging corresponding to concatenation w/o splitting. 869 if (Offset == 0 || Offset == ChainPred->Nodes.size()) 870 return; 871 // Skip merging if it breaks Forced successors. 872 NodeT *Node = ChainPred->Nodes[Offset - 1]; 873 if (Node->ForcedSucc != nullptr) 874 return; 875 // Apply the merge, compute the corresponding gain, and update the best 876 // value, if the merge is beneficial. 877 for (const MergeTypeT &MergeType : MergeTypes) { 878 Gain.updateIfLessThan( 879 computeMergeGain(ChainPred, ChainSucc, Jumps, Offset, MergeType)); 880 } 881 }; 882 883 // Try to concatenate two chains w/o splitting. 884 Gain.updateIfLessThan( 885 computeMergeGain(ChainPred, ChainSucc, Jumps, 0, MergeTypeT::X_Y)); 886 887 // Attach (a part of) ChainPred before the first node of ChainSucc. 888 for (JumpT *Jump : ChainSucc->Nodes.front()->InJumps) { 889 const NodeT *SrcBlock = Jump->Source; 890 if (SrcBlock->CurChain != ChainPred) 891 continue; 892 size_t Offset = SrcBlock->CurIndex + 1; 893 tryChainMerging(Offset, {MergeTypeT::X1_Y_X2, MergeTypeT::X2_X1_Y}); 894 } 895 896 // Attach (a part of) ChainPred after the last node of ChainSucc. 897 for (JumpT *Jump : ChainSucc->Nodes.back()->OutJumps) { 898 const NodeT *DstBlock = Jump->Target; 899 if (DstBlock->CurChain != ChainPred) 900 continue; 901 size_t Offset = DstBlock->CurIndex; 902 tryChainMerging(Offset, {MergeTypeT::X1_Y_X2, MergeTypeT::Y_X2_X1}); 903 } 904 905 // Try to break ChainPred in various ways and concatenate with ChainSucc. 906 if (ChainPred->Nodes.size() <= ChainSplitThreshold) { 907 for (size_t Offset = 1; Offset < ChainPred->Nodes.size(); Offset++) { 908 // Do not split the chain along a fall-through jump. One of the two 909 // loops above may still "break" such a jump whenever it results in a 910 // new fall-through. 911 const NodeT *BB = ChainPred->Nodes[Offset - 1]; 912 const NodeT *BB2 = ChainPred->Nodes[Offset]; 913 if (BB->isSuccessor(BB2)) 914 continue; 915 916 // In practice, applying X2_Y_X1 merging almost never provides benefits; 917 // thus, we exclude it from consideration to reduce the search space. 918 tryChainMerging(Offset, {MergeTypeT::X1_Y_X2, MergeTypeT::Y_X2_X1, 919 MergeTypeT::X2_X1_Y}); 920 } 921 } 922 923 Edge->setCachedMergeGain(ChainPred, ChainSucc, Gain); 924 return Gain; 925 } 926 927 /// Compute the score gain of merging two chains, respecting a given 928 /// merge 'type' and 'offset'. 929 /// 930 /// The two chains are not modified in the method. 931 MergeGainT computeMergeGain(const ChainT *ChainPred, const ChainT *ChainSucc, 932 const MergedJumpsT &Jumps, size_t MergeOffset, 933 MergeTypeT MergeType) const { 934 MergedNodesT MergedNodes = 935 mergeNodes(ChainPred->Nodes, ChainSucc->Nodes, MergeOffset, MergeType); 936 937 // Do not allow a merge that does not preserve the original entry point. 938 if ((ChainPred->isEntry() || ChainSucc->isEntry()) && 939 !MergedNodes.getFirstNode()->isEntry()) 940 return MergeGainT(); 941 942 // The gain for the new chain. 943 double NewScore = extTSPScore(MergedNodes, Jumps); 944 double CurScore = ChainPred->Score; 945 return MergeGainT(NewScore - CurScore, MergeOffset, MergeType); 946 } 947 948 /// Merge chain From into chain Into, update the list of active chains, 949 /// adjacency information, and the corresponding cached values. 950 void mergeChains(ChainT *Into, ChainT *From, size_t MergeOffset, 951 MergeTypeT MergeType) { 952 assert(Into != From && "a chain cannot be merged with itself"); 953 954 // Merge the nodes. 955 MergedNodesT MergedNodes = 956 mergeNodes(Into->Nodes, From->Nodes, MergeOffset, MergeType); 957 Into->merge(From, MergedNodes.getNodes()); 958 959 // Merge the edges. 960 Into->mergeEdges(From); 961 From->clear(); 962 963 // Update cached ext-tsp score for the new chain. 964 ChainEdge *SelfEdge = Into->getEdge(Into); 965 if (SelfEdge != nullptr) { 966 MergedNodes = MergedNodesT(Into->Nodes.begin(), Into->Nodes.end()); 967 MergedJumpsT MergedJumps(&SelfEdge->jumps()); 968 Into->Score = extTSPScore(MergedNodes, MergedJumps); 969 } 970 971 // Remove the chain from the list of active chains. 972 llvm::erase(HotChains, From); 973 974 // Invalidate caches. 975 for (auto EdgeIt : Into->Edges) 976 EdgeIt.second->invalidateCache(); 977 } 978 979 /// Concatenate all chains into the final order. 980 std::vector<uint64_t> concatChains() { 981 // Collect non-empty chains. 982 std::vector<const ChainT *> SortedChains; 983 for (ChainT &Chain : AllChains) { 984 if (!Chain.Nodes.empty()) 985 SortedChains.push_back(&Chain); 986 } 987 988 // Sorting chains by density in the decreasing order. 989 std::sort(SortedChains.begin(), SortedChains.end(), 990 [&](const ChainT *L, const ChainT *R) { 991 // Place the entry point at the beginning of the order. 992 if (L->isEntry() != R->isEntry()) 993 return L->isEntry(); 994 995 // Compare by density and break ties by chain identifiers. 996 return std::make_tuple(-L->density(), L->Id) < 997 std::make_tuple(-R->density(), R->Id); 998 }); 999 1000 // Collect the nodes in the order specified by their chains. 1001 std::vector<uint64_t> Order; 1002 Order.reserve(NumNodes); 1003 for (const ChainT *Chain : SortedChains) 1004 for (NodeT *Node : Chain->Nodes) 1005 Order.push_back(Node->Index); 1006 return Order; 1007 } 1008 1009 private: 1010 /// The number of nodes in the graph. 1011 const size_t NumNodes; 1012 1013 /// Successors of each node. 1014 std::vector<std::vector<uint64_t>> SuccNodes; 1015 1016 /// Predecessors of each node. 1017 std::vector<std::vector<uint64_t>> PredNodes; 1018 1019 /// All nodes (basic blocks) in the graph. 1020 std::vector<NodeT> AllNodes; 1021 1022 /// All jumps between the nodes. 1023 std::vector<JumpT> AllJumps; 1024 1025 /// All chains of nodes. 1026 std::vector<ChainT> AllChains; 1027 1028 /// All edges between the chains. 1029 std::vector<ChainEdge> AllEdges; 1030 1031 /// Active chains. The vector gets updated at runtime when chains are merged. 1032 std::vector<ChainT *> HotChains; 1033 }; 1034 1035 /// The implementation of the Cache-Directed Sort (CDSort) algorithm for 1036 /// ordering functions represented by a call graph. 1037 class CDSortImpl { 1038 public: 1039 CDSortImpl(const CDSortConfig &Config, ArrayRef<uint64_t> NodeSizes, 1040 ArrayRef<uint64_t> NodeCounts, ArrayRef<EdgeCount> EdgeCounts, 1041 ArrayRef<uint64_t> EdgeOffsets) 1042 : Config(Config), NumNodes(NodeSizes.size()) { 1043 initialize(NodeSizes, NodeCounts, EdgeCounts, EdgeOffsets); 1044 } 1045 1046 /// Run the algorithm and return an ordered set of function clusters. 1047 std::vector<uint64_t> run() { 1048 // Merge pairs of chains while improving the objective. 1049 mergeChainPairs(); 1050 1051 // Collect nodes from all the chains. 1052 return concatChains(); 1053 } 1054 1055 private: 1056 /// Initialize the algorithm's data structures. 1057 void initialize(const ArrayRef<uint64_t> &NodeSizes, 1058 const ArrayRef<uint64_t> &NodeCounts, 1059 const ArrayRef<EdgeCount> &EdgeCounts, 1060 const ArrayRef<uint64_t> &EdgeOffsets) { 1061 // Initialize nodes. 1062 AllNodes.reserve(NumNodes); 1063 for (uint64_t Node = 0; Node < NumNodes; Node++) { 1064 uint64_t Size = std::max<uint64_t>(NodeSizes[Node], 1ULL); 1065 uint64_t ExecutionCount = NodeCounts[Node]; 1066 AllNodes.emplace_back(Node, Size, ExecutionCount); 1067 TotalSamples += ExecutionCount; 1068 if (ExecutionCount > 0) 1069 TotalSize += Size; 1070 } 1071 1072 // Initialize jumps between the nodes. 1073 SuccNodes.resize(NumNodes); 1074 PredNodes.resize(NumNodes); 1075 AllJumps.reserve(EdgeCounts.size()); 1076 for (size_t I = 0; I < EdgeCounts.size(); I++) { 1077 auto [Pred, Succ, Count] = EdgeCounts[I]; 1078 // Ignore recursive calls. 1079 if (Pred == Succ) 1080 continue; 1081 1082 SuccNodes[Pred].push_back(Succ); 1083 PredNodes[Succ].push_back(Pred); 1084 if (Count > 0) { 1085 NodeT &PredNode = AllNodes[Pred]; 1086 NodeT &SuccNode = AllNodes[Succ]; 1087 AllJumps.emplace_back(&PredNode, &SuccNode, Count); 1088 AllJumps.back().Offset = EdgeOffsets[I]; 1089 SuccNode.InJumps.push_back(&AllJumps.back()); 1090 PredNode.OutJumps.push_back(&AllJumps.back()); 1091 // Adjust execution counts. 1092 PredNode.ExecutionCount = std::max(PredNode.ExecutionCount, Count); 1093 SuccNode.ExecutionCount = std::max(SuccNode.ExecutionCount, Count); 1094 } 1095 } 1096 1097 // Initialize chains. 1098 AllChains.reserve(NumNodes); 1099 for (NodeT &Node : AllNodes) { 1100 // Adjust execution counts. 1101 Node.ExecutionCount = std::max(Node.ExecutionCount, Node.inCount()); 1102 Node.ExecutionCount = std::max(Node.ExecutionCount, Node.outCount()); 1103 // Create chain. 1104 AllChains.emplace_back(Node.Index, &Node); 1105 Node.CurChain = &AllChains.back(); 1106 } 1107 1108 // Initialize chain edges. 1109 AllEdges.reserve(AllJumps.size()); 1110 for (NodeT &PredNode : AllNodes) { 1111 for (JumpT *Jump : PredNode.OutJumps) { 1112 NodeT *SuccNode = Jump->Target; 1113 ChainEdge *CurEdge = PredNode.CurChain->getEdge(SuccNode->CurChain); 1114 // This edge is already present in the graph. 1115 if (CurEdge != nullptr) { 1116 assert(SuccNode->CurChain->getEdge(PredNode.CurChain) != nullptr); 1117 CurEdge->appendJump(Jump); 1118 continue; 1119 } 1120 // This is a new edge. 1121 AllEdges.emplace_back(Jump); 1122 PredNode.CurChain->addEdge(SuccNode->CurChain, &AllEdges.back()); 1123 SuccNode->CurChain->addEdge(PredNode.CurChain, &AllEdges.back()); 1124 } 1125 } 1126 } 1127 1128 /// Merge pairs of chains while there is an improvement in the objective. 1129 void mergeChainPairs() { 1130 // Create a priority queue containing all edges ordered by the merge gain. 1131 auto GainComparator = [](ChainEdge *L, ChainEdge *R) { 1132 return std::make_tuple(-L->gain(), L->srcChain()->Id, L->dstChain()->Id) < 1133 std::make_tuple(-R->gain(), R->srcChain()->Id, R->dstChain()->Id); 1134 }; 1135 std::set<ChainEdge *, decltype(GainComparator)> Queue(GainComparator); 1136 1137 // Insert the edges into the queue. 1138 [[maybe_unused]] size_t NumActiveChains = 0; 1139 for (NodeT &Node : AllNodes) { 1140 if (Node.ExecutionCount == 0) 1141 continue; 1142 ++NumActiveChains; 1143 for (const auto &[_, Edge] : Node.CurChain->Edges) { 1144 // Ignore self-edges. 1145 if (Edge->isSelfEdge()) 1146 continue; 1147 // Ignore already processed edges. 1148 if (Edge->gain() != -1.0) 1149 continue; 1150 1151 // Compute the gain of merging the two chains. 1152 MergeGainT Gain = getBestMergeGain(Edge); 1153 Edge->setMergeGain(Gain); 1154 1155 if (Edge->gain() > EPS) 1156 Queue.insert(Edge); 1157 } 1158 } 1159 1160 // Merge the chains while the gain of merging is positive. 1161 while (!Queue.empty()) { 1162 // Extract the best (top) edge for merging. 1163 ChainEdge *BestEdge = *Queue.begin(); 1164 Queue.erase(Queue.begin()); 1165 ChainT *BestSrcChain = BestEdge->srcChain(); 1166 ChainT *BestDstChain = BestEdge->dstChain(); 1167 1168 // Remove outdated edges from the queue. 1169 for (const auto &[_, ChainEdge] : BestSrcChain->Edges) 1170 Queue.erase(ChainEdge); 1171 for (const auto &[_, ChainEdge] : BestDstChain->Edges) 1172 Queue.erase(ChainEdge); 1173 1174 // Merge the best pair of chains. 1175 MergeGainT BestGain = BestEdge->getMergeGain(); 1176 mergeChains(BestSrcChain, BestDstChain, BestGain.mergeOffset(), 1177 BestGain.mergeType()); 1178 --NumActiveChains; 1179 1180 // Insert newly created edges into the queue. 1181 for (const auto &[_, Edge] : BestSrcChain->Edges) { 1182 // Ignore loop edges. 1183 if (Edge->isSelfEdge()) 1184 continue; 1185 if (Edge->srcChain()->numBlocks() + Edge->dstChain()->numBlocks() > 1186 Config.MaxChainSize) 1187 continue; 1188 1189 // Compute the gain of merging the two chains. 1190 MergeGainT Gain = getBestMergeGain(Edge); 1191 Edge->setMergeGain(Gain); 1192 1193 if (Edge->gain() > EPS) 1194 Queue.insert(Edge); 1195 } 1196 } 1197 1198 LLVM_DEBUG(dbgs() << "Cache-directed function sorting reduced the number" 1199 << " of chains from " << NumNodes << " to " 1200 << NumActiveChains << "\n"); 1201 } 1202 1203 /// Compute the gain of merging two chains. 1204 /// 1205 /// The function considers all possible ways of merging two chains and 1206 /// computes the one having the largest increase in ExtTSP objective. The 1207 /// result is a pair with the first element being the gain and the second 1208 /// element being the corresponding merging type. 1209 MergeGainT getBestMergeGain(ChainEdge *Edge) const { 1210 assert(!Edge->jumps().empty() && "trying to merge chains w/o jumps"); 1211 // Precompute jumps between ChainPred and ChainSucc. 1212 MergedJumpsT Jumps(&Edge->jumps()); 1213 ChainT *SrcChain = Edge->srcChain(); 1214 ChainT *DstChain = Edge->dstChain(); 1215 1216 // This object holds the best currently chosen gain of merging two chains. 1217 MergeGainT Gain = MergeGainT(); 1218 1219 /// Given a list of merge types, try to merge two chains and update Gain 1220 /// with a better alternative. 1221 auto tryChainMerging = [&](const std::vector<MergeTypeT> &MergeTypes) { 1222 // Apply the merge, compute the corresponding gain, and update the best 1223 // value, if the merge is beneficial. 1224 for (const MergeTypeT &MergeType : MergeTypes) { 1225 MergeGainT NewGain = 1226 computeMergeGain(SrcChain, DstChain, Jumps, MergeType); 1227 1228 // When forward and backward gains are the same, prioritize merging that 1229 // preserves the original order of the functions in the binary. 1230 if (std::abs(Gain.score() - NewGain.score()) < EPS) { 1231 if ((MergeType == MergeTypeT::X_Y && SrcChain->Id < DstChain->Id) || 1232 (MergeType == MergeTypeT::Y_X && SrcChain->Id > DstChain->Id)) { 1233 Gain = NewGain; 1234 } 1235 } else if (NewGain.score() > Gain.score() + EPS) { 1236 Gain = NewGain; 1237 } 1238 } 1239 }; 1240 1241 // Try to concatenate two chains w/o splitting. 1242 tryChainMerging({MergeTypeT::X_Y, MergeTypeT::Y_X}); 1243 1244 return Gain; 1245 } 1246 1247 /// Compute the score gain of merging two chains, respecting a given type. 1248 /// 1249 /// The two chains are not modified in the method. 1250 MergeGainT computeMergeGain(ChainT *ChainPred, ChainT *ChainSucc, 1251 const MergedJumpsT &Jumps, 1252 MergeTypeT MergeType) const { 1253 // This doesn't depend on the ordering of the nodes 1254 double FreqGain = freqBasedLocalityGain(ChainPred, ChainSucc); 1255 1256 // Merge offset is always 0, as the chains are not split. 1257 size_t MergeOffset = 0; 1258 auto MergedBlocks = 1259 mergeNodes(ChainPred->Nodes, ChainSucc->Nodes, MergeOffset, MergeType); 1260 double DistGain = distBasedLocalityGain(MergedBlocks, Jumps); 1261 1262 double GainScore = DistGain + Config.FrequencyScale * FreqGain; 1263 // Scale the result to increase the importance of merging short chains. 1264 if (GainScore >= 0.0) 1265 GainScore /= std::min(ChainPred->Size, ChainSucc->Size); 1266 1267 return MergeGainT(GainScore, MergeOffset, MergeType); 1268 } 1269 1270 /// Compute the change of the frequency locality after merging the chains. 1271 double freqBasedLocalityGain(ChainT *ChainPred, ChainT *ChainSucc) const { 1272 auto missProbability = [&](double ChainDensity) { 1273 double PageSamples = ChainDensity * Config.CacheSize; 1274 if (PageSamples >= TotalSamples) 1275 return 0.0; 1276 double P = PageSamples / TotalSamples; 1277 return pow(1.0 - P, static_cast<double>(Config.CacheEntries)); 1278 }; 1279 1280 // Cache misses on the chains before merging. 1281 double CurScore = 1282 ChainPred->ExecutionCount * missProbability(ChainPred->density()) + 1283 ChainSucc->ExecutionCount * missProbability(ChainSucc->density()); 1284 1285 // Cache misses on the merged chain 1286 double MergedCounts = ChainPred->ExecutionCount + ChainSucc->ExecutionCount; 1287 double MergedSize = ChainPred->Size + ChainSucc->Size; 1288 double MergedDensity = static_cast<double>(MergedCounts) / MergedSize; 1289 double NewScore = MergedCounts * missProbability(MergedDensity); 1290 1291 return CurScore - NewScore; 1292 } 1293 1294 /// Compute the distance locality for a jump / call. 1295 double distScore(uint64_t SrcAddr, uint64_t DstAddr, uint64_t Count) const { 1296 uint64_t Dist = SrcAddr <= DstAddr ? DstAddr - SrcAddr : SrcAddr - DstAddr; 1297 double D = Dist == 0 ? 0.1 : static_cast<double>(Dist); 1298 return static_cast<double>(Count) * std::pow(D, -Config.DistancePower); 1299 } 1300 1301 /// Compute the change of the distance locality after merging the chains. 1302 double distBasedLocalityGain(const MergedNodesT &Nodes, 1303 const MergedJumpsT &Jumps) const { 1304 uint64_t CurAddr = 0; 1305 Nodes.forEach([&](const NodeT *Node) { 1306 Node->EstimatedAddr = CurAddr; 1307 CurAddr += Node->Size; 1308 }); 1309 1310 double CurScore = 0; 1311 double NewScore = 0; 1312 Jumps.forEach([&](const JumpT *Jump) { 1313 uint64_t SrcAddr = Jump->Source->EstimatedAddr + Jump->Offset; 1314 uint64_t DstAddr = Jump->Target->EstimatedAddr; 1315 NewScore += distScore(SrcAddr, DstAddr, Jump->ExecutionCount); 1316 CurScore += distScore(0, TotalSize, Jump->ExecutionCount); 1317 }); 1318 return NewScore - CurScore; 1319 } 1320 1321 /// Merge chain From into chain Into, update the list of active chains, 1322 /// adjacency information, and the corresponding cached values. 1323 void mergeChains(ChainT *Into, ChainT *From, size_t MergeOffset, 1324 MergeTypeT MergeType) { 1325 assert(Into != From && "a chain cannot be merged with itself"); 1326 1327 // Merge the nodes. 1328 MergedNodesT MergedNodes = 1329 mergeNodes(Into->Nodes, From->Nodes, MergeOffset, MergeType); 1330 Into->merge(From, MergedNodes.getNodes()); 1331 1332 // Merge the edges. 1333 Into->mergeEdges(From); 1334 From->clear(); 1335 } 1336 1337 /// Concatenate all chains into the final order. 1338 std::vector<uint64_t> concatChains() { 1339 // Collect chains and calculate density stats for their sorting. 1340 std::vector<const ChainT *> SortedChains; 1341 DenseMap<const ChainT *, double> ChainDensity; 1342 for (ChainT &Chain : AllChains) { 1343 if (!Chain.Nodes.empty()) { 1344 SortedChains.push_back(&Chain); 1345 // Using doubles to avoid overflow of ExecutionCounts. 1346 double Size = 0; 1347 double ExecutionCount = 0; 1348 for (NodeT *Node : Chain.Nodes) { 1349 Size += static_cast<double>(Node->Size); 1350 ExecutionCount += static_cast<double>(Node->ExecutionCount); 1351 } 1352 assert(Size > 0 && "a chain of zero size"); 1353 ChainDensity[&Chain] = ExecutionCount / Size; 1354 } 1355 } 1356 1357 // Sort chains by density in the decreasing order. 1358 std::sort(SortedChains.begin(), SortedChains.end(), 1359 [&](const ChainT *L, const ChainT *R) { 1360 const double DL = ChainDensity[L]; 1361 const double DR = ChainDensity[R]; 1362 // Compare by density and break ties by chain identifiers. 1363 return std::make_tuple(-DL, L->Id) < 1364 std::make_tuple(-DR, R->Id); 1365 }); 1366 1367 // Collect the nodes in the order specified by their chains. 1368 std::vector<uint64_t> Order; 1369 Order.reserve(NumNodes); 1370 for (const ChainT *Chain : SortedChains) 1371 for (NodeT *Node : Chain->Nodes) 1372 Order.push_back(Node->Index); 1373 return Order; 1374 } 1375 1376 private: 1377 /// Config for the algorithm. 1378 const CDSortConfig Config; 1379 1380 /// The number of nodes in the graph. 1381 const size_t NumNodes; 1382 1383 /// Successors of each node. 1384 std::vector<std::vector<uint64_t>> SuccNodes; 1385 1386 /// Predecessors of each node. 1387 std::vector<std::vector<uint64_t>> PredNodes; 1388 1389 /// All nodes (functions) in the graph. 1390 std::vector<NodeT> AllNodes; 1391 1392 /// All jumps (function calls) between the nodes. 1393 std::vector<JumpT> AllJumps; 1394 1395 /// All chains of nodes. 1396 std::vector<ChainT> AllChains; 1397 1398 /// All edges between the chains. 1399 std::vector<ChainEdge> AllEdges; 1400 1401 /// The total number of samples in the graph. 1402 uint64_t TotalSamples{0}; 1403 1404 /// The total size of the nodes in the graph. 1405 uint64_t TotalSize{0}; 1406 }; 1407 1408 } // end of anonymous namespace 1409 1410 std::vector<uint64_t> 1411 codelayout::computeExtTspLayout(ArrayRef<uint64_t> NodeSizes, 1412 ArrayRef<uint64_t> NodeCounts, 1413 ArrayRef<EdgeCount> EdgeCounts) { 1414 // Verify correctness of the input data. 1415 assert(NodeCounts.size() == NodeSizes.size() && "Incorrect input"); 1416 assert(NodeSizes.size() > 2 && "Incorrect input"); 1417 1418 // Apply the reordering algorithm. 1419 ExtTSPImpl Alg(NodeSizes, NodeCounts, EdgeCounts); 1420 std::vector<uint64_t> Result = Alg.run(); 1421 1422 // Verify correctness of the output. 1423 assert(Result.front() == 0 && "Original entry point is not preserved"); 1424 assert(Result.size() == NodeSizes.size() && "Incorrect size of layout"); 1425 return Result; 1426 } 1427 1428 double codelayout::calcExtTspScore(ArrayRef<uint64_t> Order, 1429 ArrayRef<uint64_t> NodeSizes, 1430 ArrayRef<uint64_t> NodeCounts, 1431 ArrayRef<EdgeCount> EdgeCounts) { 1432 // Estimate addresses of the blocks in memory. 1433 std::vector<uint64_t> Addr(NodeSizes.size(), 0); 1434 for (size_t Idx = 1; Idx < Order.size(); Idx++) { 1435 Addr[Order[Idx]] = Addr[Order[Idx - 1]] + NodeSizes[Order[Idx - 1]]; 1436 } 1437 std::vector<uint64_t> OutDegree(NodeSizes.size(), 0); 1438 for (auto Edge : EdgeCounts) 1439 ++OutDegree[Edge.src]; 1440 1441 // Increase the score for each jump. 1442 double Score = 0; 1443 for (auto Edge : EdgeCounts) { 1444 bool IsConditional = OutDegree[Edge.src] > 1; 1445 Score += ::extTSPScore(Addr[Edge.src], NodeSizes[Edge.src], Addr[Edge.dst], 1446 Edge.count, IsConditional); 1447 } 1448 return Score; 1449 } 1450 1451 double codelayout::calcExtTspScore(ArrayRef<uint64_t> NodeSizes, 1452 ArrayRef<uint64_t> NodeCounts, 1453 ArrayRef<EdgeCount> EdgeCounts) { 1454 std::vector<uint64_t> Order(NodeSizes.size()); 1455 for (size_t Idx = 0; Idx < NodeSizes.size(); Idx++) { 1456 Order[Idx] = Idx; 1457 } 1458 return calcExtTspScore(Order, NodeSizes, NodeCounts, EdgeCounts); 1459 } 1460 1461 std::vector<uint64_t> codelayout::computeCacheDirectedLayout( 1462 const CDSortConfig &Config, ArrayRef<uint64_t> FuncSizes, 1463 ArrayRef<uint64_t> FuncCounts, ArrayRef<EdgeCount> CallCounts, 1464 ArrayRef<uint64_t> CallOffsets) { 1465 // Verify correctness of the input data. 1466 assert(FuncCounts.size() == FuncSizes.size() && "Incorrect input"); 1467 1468 // Apply the reordering algorithm. 1469 CDSortImpl Alg(Config, FuncSizes, FuncCounts, CallCounts, CallOffsets); 1470 std::vector<uint64_t> Result = Alg.run(); 1471 assert(Result.size() == FuncSizes.size() && "Incorrect size of layout"); 1472 return Result; 1473 } 1474 1475 std::vector<uint64_t> codelayout::computeCacheDirectedLayout( 1476 ArrayRef<uint64_t> FuncSizes, ArrayRef<uint64_t> FuncCounts, 1477 ArrayRef<EdgeCount> CallCounts, ArrayRef<uint64_t> CallOffsets) { 1478 CDSortConfig Config; 1479 // Populate the config from the command-line options. 1480 if (CacheEntries.getNumOccurrences() > 0) 1481 Config.CacheEntries = CacheEntries; 1482 if (CacheSize.getNumOccurrences() > 0) 1483 Config.CacheSize = CacheSize; 1484 if (CDMaxChainSize.getNumOccurrences() > 0) 1485 Config.MaxChainSize = CDMaxChainSize; 1486 if (DistancePower.getNumOccurrences() > 0) 1487 Config.DistancePower = DistancePower; 1488 if (FrequencyScale.getNumOccurrences() > 0) 1489 Config.FrequencyScale = FrequencyScale; 1490 return computeCacheDirectedLayout(Config, FuncSizes, FuncCounts, CallCounts, 1491 CallOffsets); 1492 } 1493