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 // ExtTSP - layout of basic blocks with i-cache optimization. 10 // 11 // The algorithm tries to find a layout of nodes (basic blocks) of a given CFG 12 // optimizing jump locality and thus processor I-cache utilization. This is 13 // achieved via increasing the number of fall-through jumps and co-locating 14 // frequently executed nodes together. The name follows the underlying 15 // optimization problem, Extended-TSP, which is a generalization of classical 16 // (maximum) Traveling Salesmen Problem. 17 // 18 // The algorithm is a greedy heuristic that works with chains (ordered lists) 19 // of basic blocks. Initially all chains are isolated basic blocks. On every 20 // iteration, we pick a pair of chains whose merging yields the biggest increase 21 // in the ExtTSP score, which models how i-cache "friendly" a specific chain is. 22 // A pair of chains giving the maximum gain is merged into a new chain. The 23 // procedure stops when there is only one chain left, or when merging does not 24 // increase ExtTSP. In the latter case, the remaining chains are sorted by 25 // density in the decreasing order. 26 // 27 // An important aspect is the way two chains are merged. Unlike earlier 28 // algorithms (e.g., based on the approach of Pettis-Hansen), two 29 // chains, X and Y, are first split into three, X1, X2, and Y. Then we 30 // consider all possible ways of gluing the three chains (e.g., X1YX2, X1X2Y, 31 // X2X1Y, X2YX1, YX1X2, YX2X1) and choose the one producing the largest score. 32 // This improves the quality of the final result (the search space is larger) 33 // while keeping the implementation sufficiently fast. 34 // 35 // Reference: 36 // * A. Newell and S. Pupyrev, Improved Basic Block Reordering, 37 // IEEE Transactions on Computers, 2020 38 // 39 //===----------------------------------------------------------------------===// 40 41 #include "llvm/Transforms/Utils/CodeLayout.h" 42 #include "llvm/Support/CommandLine.h" 43 44 using namespace llvm; 45 #define DEBUG_TYPE "code-layout" 46 47 cl::opt<bool> EnableExtTspBlockPlacement( 48 "enable-ext-tsp-block-placement", cl::Hidden, cl::init(false), 49 cl::desc("Enable machine block placement based on the ext-tsp model, " 50 "optimizing I-cache utilization.")); 51 52 cl::opt<bool> ApplyExtTspWithoutProfile( 53 "ext-tsp-apply-without-profile", 54 cl::desc("Whether to apply ext-tsp placement for instances w/o profile"), 55 cl::init(true), cl::Hidden); 56 57 // Algorithm-specific constants. The values are tuned for the best performance 58 // of large-scale front-end bound binaries. 59 static cl::opt<double> 60 ForwardWeight("ext-tsp-forward-weight", cl::Hidden, cl::init(0.1), 61 cl::desc("The weight of forward jumps for ExtTSP value")); 62 63 static cl::opt<double> 64 BackwardWeight("ext-tsp-backward-weight", cl::Hidden, cl::init(0.1), 65 cl::desc("The weight of backward jumps for ExtTSP value")); 66 67 static cl::opt<unsigned> ForwardDistance( 68 "ext-tsp-forward-distance", cl::Hidden, cl::init(1024), 69 cl::desc("The maximum distance (in bytes) of a forward jump for ExtTSP")); 70 71 static cl::opt<unsigned> BackwardDistance( 72 "ext-tsp-backward-distance", cl::Hidden, cl::init(640), 73 cl::desc("The maximum distance (in bytes) of a backward jump for ExtTSP")); 74 75 // The maximum size of a chain created by the algorithm. The size is bounded 76 // so that the algorithm can efficiently process extremely large instance. 77 static cl::opt<unsigned> 78 MaxChainSize("ext-tsp-max-chain-size", cl::Hidden, cl::init(4096), 79 cl::desc("The maximum size of a chain to create.")); 80 81 // The maximum size of a chain for splitting. Larger values of the threshold 82 // may yield better quality at the cost of worsen run-time. 83 static cl::opt<unsigned> ChainSplitThreshold( 84 "ext-tsp-chain-split-threshold", cl::Hidden, cl::init(128), 85 cl::desc("The maximum size of a chain to apply splitting")); 86 87 // The option enables splitting (large) chains along in-coming and out-going 88 // jumps. This typically results in a better quality. 89 static cl::opt<bool> EnableChainSplitAlongJumps( 90 "ext-tsp-enable-chain-split-along-jumps", cl::Hidden, cl::init(true), 91 cl::desc("The maximum size of a chain to apply splitting")); 92 93 namespace { 94 95 // Epsilon for comparison of doubles. 96 constexpr double EPS = 1e-8; 97 98 // Compute the Ext-TSP score for a jump between a given pair of blocks, 99 // using their sizes, (estimated) addresses and the jump execution count. 100 double extTSPScore(uint64_t SrcAddr, uint64_t SrcSize, uint64_t DstAddr, 101 uint64_t Count) { 102 // Fallthrough 103 if (SrcAddr + SrcSize == DstAddr) { 104 // Assume that FallthroughWeight = 1.0 after normalization 105 return static_cast<double>(Count); 106 } 107 // Forward 108 if (SrcAddr + SrcSize < DstAddr) { 109 const auto Dist = DstAddr - (SrcAddr + SrcSize); 110 if (Dist <= ForwardDistance) { 111 double Prob = 1.0 - static_cast<double>(Dist) / ForwardDistance; 112 return ForwardWeight * Prob * Count; 113 } 114 return 0; 115 } 116 // Backward 117 const auto Dist = SrcAddr + SrcSize - DstAddr; 118 if (Dist <= BackwardDistance) { 119 double Prob = 1.0 - static_cast<double>(Dist) / BackwardDistance; 120 return BackwardWeight * Prob * Count; 121 } 122 return 0; 123 } 124 125 /// A type of merging two chains, X and Y. The former chain is split into 126 /// X1 and X2 and then concatenated with Y in the order specified by the type. 127 enum class MergeTypeTy : int { X_Y, X1_Y_X2, Y_X2_X1, X2_X1_Y }; 128 129 /// The gain of merging two chains, that is, the Ext-TSP score of the merge 130 /// together with the corresponfiding merge 'type' and 'offset'. 131 class MergeGainTy { 132 public: 133 explicit MergeGainTy() = default; 134 explicit MergeGainTy(double Score, size_t MergeOffset, MergeTypeTy MergeType) 135 : Score(Score), MergeOffset(MergeOffset), MergeType(MergeType) {} 136 137 double score() const { return Score; } 138 139 size_t mergeOffset() const { return MergeOffset; } 140 141 MergeTypeTy mergeType() const { return MergeType; } 142 143 // Returns 'true' iff Other is preferred over this. 144 bool operator<(const MergeGainTy &Other) const { 145 return (Other.Score > EPS && Other.Score > Score + EPS); 146 } 147 148 // Update the current gain if Other is preferred over this. 149 void updateIfLessThan(const MergeGainTy &Other) { 150 if (*this < Other) 151 *this = Other; 152 } 153 154 private: 155 double Score{-1.0}; 156 size_t MergeOffset{0}; 157 MergeTypeTy MergeType{MergeTypeTy::X_Y}; 158 }; 159 160 class Jump; 161 class Chain; 162 class ChainEdge; 163 164 /// A node in the graph, typically corresponding to a basic block in CFG. 165 class Block { 166 public: 167 Block(const Block &) = delete; 168 Block(Block &&) = default; 169 Block &operator=(const Block &) = delete; 170 Block &operator=(Block &&) = default; 171 172 // The original index of the block in CFG. 173 size_t Index{0}; 174 // The index of the block in the current chain. 175 size_t CurIndex{0}; 176 // Size of the block in the binary. 177 uint64_t Size{0}; 178 // Execution count of the block in the profile data. 179 uint64_t ExecutionCount{0}; 180 // Current chain of the node. 181 Chain *CurChain{nullptr}; 182 // An offset of the block in the current chain. 183 mutable uint64_t EstimatedAddr{0}; 184 // Forced successor of the block in CFG. 185 Block *ForcedSucc{nullptr}; 186 // Forced predecessor of the block in CFG. 187 Block *ForcedPred{nullptr}; 188 // Outgoing jumps from the block. 189 std::vector<Jump *> OutJumps; 190 // Incoming jumps to the block. 191 std::vector<Jump *> InJumps; 192 193 public: 194 explicit Block(size_t Index, uint64_t Size_, uint64_t EC) 195 : Index(Index), Size(Size_), ExecutionCount(EC) {} 196 bool isEntry() const { return Index == 0; } 197 }; 198 199 /// An arc in the graph, typically corresponding to a jump between two blocks. 200 class Jump { 201 public: 202 Jump(const Jump &) = delete; 203 Jump(Jump &&) = default; 204 Jump &operator=(const Jump &) = delete; 205 Jump &operator=(Jump &&) = default; 206 207 // Source block of the jump. 208 Block *Source; 209 // Target block of the jump. 210 Block *Target; 211 // Execution count of the arc in the profile data. 212 uint64_t ExecutionCount{0}; 213 214 public: 215 explicit Jump(Block *Source, Block *Target, uint64_t ExecutionCount) 216 : Source(Source), Target(Target), ExecutionCount(ExecutionCount) {} 217 }; 218 219 /// A chain (ordered sequence) of blocks. 220 class Chain { 221 public: 222 Chain(const Chain &) = delete; 223 Chain(Chain &&) = default; 224 Chain &operator=(const Chain &) = delete; 225 Chain &operator=(Chain &&) = default; 226 227 explicit Chain(uint64_t Id, Block *Block) 228 : Id(Id), Score(0), Blocks(1, Block) {} 229 230 uint64_t id() const { return Id; } 231 232 bool isEntry() const { return Blocks[0]->Index == 0; } 233 234 double score() const { return Score; } 235 236 void setScore(double NewScore) { Score = NewScore; } 237 238 const std::vector<Block *> &blocks() const { return Blocks; } 239 240 size_t numBlocks() const { return Blocks.size(); } 241 242 const std::vector<std::pair<Chain *, ChainEdge *>> &edges() const { 243 return Edges; 244 } 245 246 ChainEdge *getEdge(Chain *Other) const { 247 for (auto It : Edges) { 248 if (It.first == Other) 249 return It.second; 250 } 251 return nullptr; 252 } 253 254 void removeEdge(Chain *Other) { 255 auto It = Edges.begin(); 256 while (It != Edges.end()) { 257 if (It->first == Other) { 258 Edges.erase(It); 259 return; 260 } 261 It++; 262 } 263 } 264 265 void addEdge(Chain *Other, ChainEdge *Edge) { 266 Edges.push_back(std::make_pair(Other, Edge)); 267 } 268 269 void merge(Chain *Other, const std::vector<Block *> &MergedBlocks) { 270 Blocks = MergedBlocks; 271 // Update the block's chains 272 for (size_t Idx = 0; Idx < Blocks.size(); Idx++) { 273 Blocks[Idx]->CurChain = this; 274 Blocks[Idx]->CurIndex = Idx; 275 } 276 } 277 278 void mergeEdges(Chain *Other); 279 280 void clear() { 281 Blocks.clear(); 282 Blocks.shrink_to_fit(); 283 Edges.clear(); 284 Edges.shrink_to_fit(); 285 } 286 287 private: 288 // Unique chain identifier. 289 uint64_t Id; 290 // Cached ext-tsp score for the chain. 291 double Score; 292 // Blocks of the chain. 293 std::vector<Block *> Blocks; 294 // Adjacent chains and corresponding edges (lists of jumps). 295 std::vector<std::pair<Chain *, ChainEdge *>> Edges; 296 }; 297 298 /// An edge in CFG representing jumps between two chains. 299 /// When blocks are merged into chains, the edges are combined too so that 300 /// there is always at most one edge between a pair of chains 301 class ChainEdge { 302 public: 303 ChainEdge(const ChainEdge &) = delete; 304 ChainEdge(ChainEdge &&) = default; 305 ChainEdge &operator=(const ChainEdge &) = delete; 306 ChainEdge &operator=(ChainEdge &&) = default; 307 308 explicit ChainEdge(Jump *Jump) 309 : SrcChain(Jump->Source->CurChain), DstChain(Jump->Target->CurChain), 310 Jumps(1, Jump) {} 311 312 const std::vector<Jump *> &jumps() const { return Jumps; } 313 314 void changeEndpoint(Chain *From, Chain *To) { 315 if (From == SrcChain) 316 SrcChain = To; 317 if (From == DstChain) 318 DstChain = To; 319 } 320 321 void appendJump(Jump *Jump) { Jumps.push_back(Jump); } 322 323 void moveJumps(ChainEdge *Other) { 324 Jumps.insert(Jumps.end(), Other->Jumps.begin(), Other->Jumps.end()); 325 Other->Jumps.clear(); 326 Other->Jumps.shrink_to_fit(); 327 } 328 329 bool hasCachedMergeGain(Chain *Src, Chain *Dst) const { 330 return Src == SrcChain ? CacheValidForward : CacheValidBackward; 331 } 332 333 MergeGainTy getCachedMergeGain(Chain *Src, Chain *Dst) const { 334 return Src == SrcChain ? CachedGainForward : CachedGainBackward; 335 } 336 337 void setCachedMergeGain(Chain *Src, Chain *Dst, MergeGainTy MergeGain) { 338 if (Src == SrcChain) { 339 CachedGainForward = MergeGain; 340 CacheValidForward = true; 341 } else { 342 CachedGainBackward = MergeGain; 343 CacheValidBackward = true; 344 } 345 } 346 347 void invalidateCache() { 348 CacheValidForward = false; 349 CacheValidBackward = false; 350 } 351 352 private: 353 // Source chain. 354 Chain *SrcChain{nullptr}; 355 // Destination chain. 356 Chain *DstChain{nullptr}; 357 // Original jumps in the binary with correspinding execution counts. 358 std::vector<Jump *> Jumps; 359 // Cached ext-tsp value for merging the pair of chains. 360 // Since the gain of merging (Src, Dst) and (Dst, Src) might be different, 361 // we store both values here. 362 MergeGainTy CachedGainForward; 363 MergeGainTy CachedGainBackward; 364 // Whether the cached value must be recomputed. 365 bool CacheValidForward{false}; 366 bool CacheValidBackward{false}; 367 }; 368 369 void Chain::mergeEdges(Chain *Other) { 370 assert(this != Other && "cannot merge a chain with itself"); 371 372 // Update edges adjacent to chain Other 373 for (auto EdgeIt : Other->Edges) { 374 const auto DstChain = EdgeIt.first; 375 const auto DstEdge = EdgeIt.second; 376 const auto TargetChain = DstChain == Other ? this : DstChain; 377 auto CurEdge = getEdge(TargetChain); 378 if (CurEdge == nullptr) { 379 DstEdge->changeEndpoint(Other, this); 380 this->addEdge(TargetChain, DstEdge); 381 if (DstChain != this && DstChain != Other) { 382 DstChain->addEdge(this, DstEdge); 383 } 384 } else { 385 CurEdge->moveJumps(DstEdge); 386 } 387 // Cleanup leftover edge 388 if (DstChain != Other) { 389 DstChain->removeEdge(Other); 390 } 391 } 392 } 393 394 using BlockIter = std::vector<Block *>::const_iterator; 395 396 /// A wrapper around three chains of blocks; it is used to avoid extra 397 /// instantiation of the vectors. 398 class MergedChain { 399 public: 400 MergedChain(BlockIter Begin1, BlockIter End1, BlockIter Begin2 = BlockIter(), 401 BlockIter End2 = BlockIter(), BlockIter Begin3 = BlockIter(), 402 BlockIter End3 = BlockIter()) 403 : Begin1(Begin1), End1(End1), Begin2(Begin2), End2(End2), Begin3(Begin3), 404 End3(End3) {} 405 406 template <typename F> void forEach(const F &Func) const { 407 for (auto It = Begin1; It != End1; It++) 408 Func(*It); 409 for (auto It = Begin2; It != End2; It++) 410 Func(*It); 411 for (auto It = Begin3; It != End3; It++) 412 Func(*It); 413 } 414 415 std::vector<Block *> getBlocks() const { 416 std::vector<Block *> Result; 417 Result.reserve(std::distance(Begin1, End1) + std::distance(Begin2, End2) + 418 std::distance(Begin3, End3)); 419 Result.insert(Result.end(), Begin1, End1); 420 Result.insert(Result.end(), Begin2, End2); 421 Result.insert(Result.end(), Begin3, End3); 422 return Result; 423 } 424 425 const Block *getFirstBlock() const { return *Begin1; } 426 427 private: 428 BlockIter Begin1; 429 BlockIter End1; 430 BlockIter Begin2; 431 BlockIter End2; 432 BlockIter Begin3; 433 BlockIter End3; 434 }; 435 436 /// The implementation of the ExtTSP algorithm. 437 class ExtTSPImpl { 438 using EdgeT = std::pair<uint64_t, uint64_t>; 439 using EdgeCountMap = DenseMap<EdgeT, uint64_t>; 440 441 public: 442 ExtTSPImpl(size_t NumNodes, const std::vector<uint64_t> &NodeSizes, 443 const std::vector<uint64_t> &NodeCounts, 444 const EdgeCountMap &EdgeCounts) 445 : NumNodes(NumNodes) { 446 initialize(NodeSizes, NodeCounts, EdgeCounts); 447 } 448 449 /// Run the algorithm and return an optimized ordering of blocks. 450 void run(std::vector<uint64_t> &Result) { 451 // Pass 1: Merge blocks with their mutually forced successors 452 mergeForcedPairs(); 453 454 // Pass 2: Merge pairs of chains while improving the ExtTSP objective 455 mergeChainPairs(); 456 457 // Pass 3: Merge cold blocks to reduce code size 458 mergeColdChains(); 459 460 // Collect blocks from all chains 461 concatChains(Result); 462 } 463 464 private: 465 /// Initialize the algorithm's data structures. 466 void initialize(const std::vector<uint64_t> &NodeSizes, 467 const std::vector<uint64_t> &NodeCounts, 468 const EdgeCountMap &EdgeCounts) { 469 // Initialize blocks 470 AllBlocks.reserve(NumNodes); 471 for (uint64_t Node = 0; Node < NumNodes; Node++) { 472 uint64_t Size = std::max<uint64_t>(NodeSizes[Node], 1ULL); 473 uint64_t ExecutionCount = NodeCounts[Node]; 474 // The execution count of the entry block is set to at least 1 475 if (Node == 0 && ExecutionCount == 0) 476 ExecutionCount = 1; 477 AllBlocks.emplace_back(Node, Size, ExecutionCount); 478 } 479 480 // Initialize jumps between blocks 481 SuccNodes = std::vector<std::vector<uint64_t>>(NumNodes); 482 PredNodes = std::vector<std::vector<uint64_t>>(NumNodes); 483 AllJumps.reserve(EdgeCounts.size()); 484 for (auto It : EdgeCounts) { 485 auto Pred = It.first.first; 486 auto Succ = It.first.second; 487 // Ignore self-edges 488 if (Pred == Succ) 489 continue; 490 491 SuccNodes[Pred].push_back(Succ); 492 PredNodes[Succ].push_back(Pred); 493 auto ExecutionCount = It.second; 494 if (ExecutionCount > 0) { 495 auto &Block = AllBlocks[Pred]; 496 auto &SuccBlock = AllBlocks[Succ]; 497 AllJumps.emplace_back(&Block, &SuccBlock, ExecutionCount); 498 SuccBlock.InJumps.push_back(&AllJumps.back()); 499 Block.OutJumps.push_back(&AllJumps.back()); 500 } 501 } 502 503 // Initialize chains 504 AllChains.reserve(NumNodes); 505 HotChains.reserve(NumNodes); 506 for (auto &Block : AllBlocks) { 507 AllChains.emplace_back(Block.Index, &Block); 508 Block.CurChain = &AllChains.back(); 509 if (Block.ExecutionCount > 0) { 510 HotChains.push_back(&AllChains.back()); 511 } 512 } 513 514 // Initialize chain edges 515 AllEdges.reserve(AllJumps.size()); 516 for (auto &Block : AllBlocks) { 517 for (auto &Jump : Block.OutJumps) { 518 auto SuccBlock = Jump->Target; 519 auto CurEdge = Block.CurChain->getEdge(SuccBlock->CurChain); 520 // this edge is already present in the graph 521 if (CurEdge != nullptr) { 522 assert(SuccBlock->CurChain->getEdge(Block.CurChain) != nullptr); 523 CurEdge->appendJump(Jump); 524 continue; 525 } 526 // this is a new edge 527 AllEdges.emplace_back(Jump); 528 Block.CurChain->addEdge(SuccBlock->CurChain, &AllEdges.back()); 529 SuccBlock->CurChain->addEdge(Block.CurChain, &AllEdges.back()); 530 } 531 } 532 } 533 534 /// For a pair of blocks, A and B, block B is the forced successor of A, 535 /// if (i) all jumps (based on profile) from A goes to B and (ii) all jumps 536 /// to B are from A. Such blocks should be adjacent in the optimal ordering; 537 /// the method finds and merges such pairs of blocks. 538 void mergeForcedPairs() { 539 // Find fallthroughs based on edge weights 540 for (auto &Block : AllBlocks) { 541 if (SuccNodes[Block.Index].size() == 1 && 542 PredNodes[SuccNodes[Block.Index][0]].size() == 1 && 543 SuccNodes[Block.Index][0] != 0) { 544 size_t SuccIndex = SuccNodes[Block.Index][0]; 545 Block.ForcedSucc = &AllBlocks[SuccIndex]; 546 AllBlocks[SuccIndex].ForcedPred = &Block; 547 } 548 } 549 550 // There might be 'cycles' in the forced dependencies, since profile 551 // data isn't 100% accurate. Typically this is observed in loops, when the 552 // loop edges are the hottest successors for the basic blocks of the loop. 553 // Break the cycles by choosing the block with the smallest index as the 554 // head. This helps to keep the original order of the loops, which likely 555 // have already been rotated in the optimized manner. 556 for (auto &Block : AllBlocks) { 557 if (Block.ForcedSucc == nullptr || Block.ForcedPred == nullptr) 558 continue; 559 560 auto SuccBlock = Block.ForcedSucc; 561 while (SuccBlock != nullptr && SuccBlock != &Block) { 562 SuccBlock = SuccBlock->ForcedSucc; 563 } 564 if (SuccBlock == nullptr) 565 continue; 566 // Break the cycle 567 AllBlocks[Block.ForcedPred->Index].ForcedSucc = nullptr; 568 Block.ForcedPred = nullptr; 569 } 570 571 // Merge blocks with their fallthrough successors 572 for (auto &Block : AllBlocks) { 573 if (Block.ForcedPred == nullptr && Block.ForcedSucc != nullptr) { 574 auto CurBlock = &Block; 575 while (CurBlock->ForcedSucc != nullptr) { 576 const auto NextBlock = CurBlock->ForcedSucc; 577 mergeChains(Block.CurChain, NextBlock->CurChain, 0, MergeTypeTy::X_Y); 578 CurBlock = NextBlock; 579 } 580 } 581 } 582 } 583 584 /// Merge pairs of chains while improving the ExtTSP objective. 585 void mergeChainPairs() { 586 /// Deterministically compare pairs of chains 587 auto compareChainPairs = [](const Chain *A1, const Chain *B1, 588 const Chain *A2, const Chain *B2) { 589 if (A1 != A2) 590 return A1->id() < A2->id(); 591 return B1->id() < B2->id(); 592 }; 593 594 while (HotChains.size() > 1) { 595 Chain *BestChainPred = nullptr; 596 Chain *BestChainSucc = nullptr; 597 auto BestGain = MergeGainTy(); 598 // Iterate over all pairs of chains 599 for (auto ChainPred : HotChains) { 600 // Get candidates for merging with the current chain 601 for (auto EdgeIter : ChainPred->edges()) { 602 auto ChainSucc = EdgeIter.first; 603 auto ChainEdge = EdgeIter.second; 604 // Ignore loop edges 605 if (ChainPred == ChainSucc) 606 continue; 607 608 // Stop early if the combined chain violates the maximum allowed size 609 if (ChainPred->numBlocks() + ChainSucc->numBlocks() >= MaxChainSize) 610 continue; 611 612 // Compute the gain of merging the two chains 613 auto CurGain = getBestMergeGain(ChainPred, ChainSucc, ChainEdge); 614 if (CurGain.score() <= EPS) 615 continue; 616 617 if (BestGain < CurGain || 618 (std::abs(CurGain.score() - BestGain.score()) < EPS && 619 compareChainPairs(ChainPred, ChainSucc, BestChainPred, 620 BestChainSucc))) { 621 BestGain = CurGain; 622 BestChainPred = ChainPred; 623 BestChainSucc = ChainSucc; 624 } 625 } 626 } 627 628 // Stop merging when there is no improvement 629 if (BestGain.score() <= EPS) 630 break; 631 632 // Merge the best pair of chains 633 mergeChains(BestChainPred, BestChainSucc, BestGain.mergeOffset(), 634 BestGain.mergeType()); 635 } 636 } 637 638 /// Merge cold blocks to reduce code size. 639 void mergeColdChains() { 640 for (size_t SrcBB = 0; SrcBB < NumNodes; SrcBB++) { 641 // Iterating over neighbors in the reverse order to make sure original 642 // fallthrough jumps are merged first 643 size_t NumSuccs = SuccNodes[SrcBB].size(); 644 for (size_t Idx = 0; Idx < NumSuccs; Idx++) { 645 auto DstBB = SuccNodes[SrcBB][NumSuccs - Idx - 1]; 646 auto SrcChain = AllBlocks[SrcBB].CurChain; 647 auto DstChain = AllBlocks[DstBB].CurChain; 648 if (SrcChain != DstChain && !DstChain->isEntry() && 649 SrcChain->blocks().back()->Index == SrcBB && 650 DstChain->blocks().front()->Index == DstBB) { 651 mergeChains(SrcChain, DstChain, 0, MergeTypeTy::X_Y); 652 } 653 } 654 } 655 } 656 657 /// Compute the Ext-TSP score for a given block order and a list of jumps. 658 double extTSPScore(const MergedChain &MergedBlocks, 659 const std::vector<Jump *> &Jumps) const { 660 if (Jumps.empty()) 661 return 0.0; 662 uint64_t CurAddr = 0; 663 MergedBlocks.forEach([&](const Block *BB) { 664 BB->EstimatedAddr = CurAddr; 665 CurAddr += BB->Size; 666 }); 667 668 double Score = 0; 669 for (auto &Jump : Jumps) { 670 const auto SrcBlock = Jump->Source; 671 const auto DstBlock = Jump->Target; 672 Score += ::extTSPScore(SrcBlock->EstimatedAddr, SrcBlock->Size, 673 DstBlock->EstimatedAddr, Jump->ExecutionCount); 674 } 675 return Score; 676 } 677 678 /// Compute the gain of merging two chains. 679 /// 680 /// The function considers all possible ways of merging two chains and 681 /// computes the one having the largest increase in ExtTSP objective. The 682 /// result is a pair with the first element being the gain and the second 683 /// element being the corresponding merging type. 684 MergeGainTy getBestMergeGain(Chain *ChainPred, Chain *ChainSucc, 685 ChainEdge *Edge) const { 686 if (Edge->hasCachedMergeGain(ChainPred, ChainSucc)) { 687 return Edge->getCachedMergeGain(ChainPred, ChainSucc); 688 } 689 690 // Precompute jumps between ChainPred and ChainSucc 691 auto Jumps = Edge->jumps(); 692 auto EdgePP = ChainPred->getEdge(ChainPred); 693 if (EdgePP != nullptr) { 694 Jumps.insert(Jumps.end(), EdgePP->jumps().begin(), EdgePP->jumps().end()); 695 } 696 assert(!Jumps.empty() && "trying to merge chains w/o jumps"); 697 698 // The object holds the best currently chosen gain of merging the two chains 699 MergeGainTy Gain = MergeGainTy(); 700 701 /// Given a merge offset and a list of merge types, try to merge two chains 702 /// and update Gain with a better alternative 703 auto tryChainMerging = [&](size_t Offset, 704 const std::vector<MergeTypeTy> &MergeTypes) { 705 // Skip merging corresponding to concatenation w/o splitting 706 if (Offset == 0 || Offset == ChainPred->blocks().size()) 707 return; 708 // Skip merging if it breaks Forced successors 709 auto BB = ChainPred->blocks()[Offset - 1]; 710 if (BB->ForcedSucc != nullptr) 711 return; 712 // Apply the merge, compute the corresponding gain, and update the best 713 // value, if the merge is beneficial 714 for (auto &MergeType : MergeTypes) { 715 Gain.updateIfLessThan( 716 computeMergeGain(ChainPred, ChainSucc, Jumps, Offset, MergeType)); 717 } 718 }; 719 720 // Try to concatenate two chains w/o splitting 721 Gain.updateIfLessThan( 722 computeMergeGain(ChainPred, ChainSucc, Jumps, 0, MergeTypeTy::X_Y)); 723 724 if (EnableChainSplitAlongJumps) { 725 // Attach (a part of) ChainPred before the first block of ChainSucc 726 for (auto &Jump : ChainSucc->blocks().front()->InJumps) { 727 const auto SrcBlock = Jump->Source; 728 if (SrcBlock->CurChain != ChainPred) 729 continue; 730 size_t Offset = SrcBlock->CurIndex + 1; 731 tryChainMerging(Offset, {MergeTypeTy::X1_Y_X2, MergeTypeTy::X2_X1_Y}); 732 } 733 734 // Attach (a part of) ChainPred after the last block of ChainSucc 735 for (auto &Jump : ChainSucc->blocks().back()->OutJumps) { 736 const auto DstBlock = Jump->Source; 737 if (DstBlock->CurChain != ChainPred) 738 continue; 739 size_t Offset = DstBlock->CurIndex; 740 tryChainMerging(Offset, {MergeTypeTy::X1_Y_X2, MergeTypeTy::Y_X2_X1}); 741 } 742 } 743 744 // Try to break ChainPred in various ways and concatenate with ChainSucc 745 if (ChainPred->blocks().size() <= ChainSplitThreshold) { 746 for (size_t Offset = 1; Offset < ChainPred->blocks().size(); Offset++) { 747 // Try to split the chain in different ways. In practice, applying 748 // X2_Y_X1 merging is almost never provides benefits; thus, we exclude 749 // it from consideration to reduce the search space 750 tryChainMerging(Offset, {MergeTypeTy::X1_Y_X2, MergeTypeTy::Y_X2_X1, 751 MergeTypeTy::X2_X1_Y}); 752 } 753 } 754 Edge->setCachedMergeGain(ChainPred, ChainSucc, Gain); 755 return Gain; 756 } 757 758 /// Compute the score gain of merging two chains, respecting a given 759 /// merge 'type' and 'offset'. 760 /// 761 /// The two chains are not modified in the method. 762 MergeGainTy computeMergeGain(const Chain *ChainPred, const Chain *ChainSucc, 763 const std::vector<Jump *> &Jumps, 764 size_t MergeOffset, 765 MergeTypeTy MergeType) const { 766 auto MergedBlocks = mergeBlocks(ChainPred->blocks(), ChainSucc->blocks(), 767 MergeOffset, MergeType); 768 769 // Do not allow a merge that does not preserve the original entry block 770 if ((ChainPred->isEntry() || ChainSucc->isEntry()) && 771 !MergedBlocks.getFirstBlock()->isEntry()) 772 return MergeGainTy(); 773 774 // The gain for the new chain 775 auto NewGainScore = extTSPScore(MergedBlocks, Jumps) - ChainPred->score(); 776 return MergeGainTy(NewGainScore, MergeOffset, MergeType); 777 } 778 779 /// Merge two chains of blocks respecting a given merge 'type' and 'offset'. 780 /// 781 /// If MergeType == 0, then the result is a concatentation of two chains. 782 /// Otherwise, the first chain is cut into two sub-chains at the offset, 783 /// and merged using all possible ways of concatenating three chains. 784 MergedChain mergeBlocks(const std::vector<Block *> &X, 785 const std::vector<Block *> &Y, size_t MergeOffset, 786 MergeTypeTy MergeType) const { 787 // Split the first chain, X, into X1 and X2 788 BlockIter BeginX1 = X.begin(); 789 BlockIter EndX1 = X.begin() + MergeOffset; 790 BlockIter BeginX2 = X.begin() + MergeOffset; 791 BlockIter EndX2 = X.end(); 792 BlockIter BeginY = Y.begin(); 793 BlockIter EndY = Y.end(); 794 795 // Construct a new chain from the three existing ones 796 switch (MergeType) { 797 case MergeTypeTy::X_Y: 798 return MergedChain(BeginX1, EndX2, BeginY, EndY); 799 case MergeTypeTy::X1_Y_X2: 800 return MergedChain(BeginX1, EndX1, BeginY, EndY, BeginX2, EndX2); 801 case MergeTypeTy::Y_X2_X1: 802 return MergedChain(BeginY, EndY, BeginX2, EndX2, BeginX1, EndX1); 803 case MergeTypeTy::X2_X1_Y: 804 return MergedChain(BeginX2, EndX2, BeginX1, EndX1, BeginY, EndY); 805 } 806 llvm_unreachable("unexpected chain merge type"); 807 } 808 809 /// Merge chain From into chain Into, update the list of active chains, 810 /// adjacency information, and the corresponding cached values. 811 void mergeChains(Chain *Into, Chain *From, size_t MergeOffset, 812 MergeTypeTy MergeType) { 813 assert(Into != From && "a chain cannot be merged with itself"); 814 815 // Merge the blocks 816 auto MergedBlocks = 817 mergeBlocks(Into->blocks(), From->blocks(), MergeOffset, MergeType); 818 Into->merge(From, MergedBlocks.getBlocks()); 819 Into->mergeEdges(From); 820 From->clear(); 821 822 // Update cached ext-tsp score for the new chain 823 auto SelfEdge = Into->getEdge(Into); 824 if (SelfEdge != nullptr) { 825 MergedBlocks = MergedChain(Into->blocks().begin(), Into->blocks().end()); 826 Into->setScore(extTSPScore(MergedBlocks, SelfEdge->jumps())); 827 } 828 829 // Remove chain From from the list of active chains 830 auto Iter = std::remove(HotChains.begin(), HotChains.end(), From); 831 HotChains.erase(Iter, HotChains.end()); 832 833 // Invalidate caches 834 for (auto EdgeIter : Into->edges()) { 835 EdgeIter.second->invalidateCache(); 836 } 837 } 838 839 /// Concatenate all chains into a final order of blocks. 840 void concatChains(std::vector<uint64_t> &Order) { 841 // Collect chains and calculate some stats for their sorting 842 std::vector<Chain *> SortedChains; 843 DenseMap<const Chain *, double> ChainDensity; 844 for (auto &Chain : AllChains) { 845 if (!Chain.blocks().empty()) { 846 SortedChains.push_back(&Chain); 847 // Using doubles to avoid overflow of ExecutionCount 848 double Size = 0; 849 double ExecutionCount = 0; 850 for (auto Block : Chain.blocks()) { 851 Size += static_cast<double>(Block->Size); 852 ExecutionCount += static_cast<double>(Block->ExecutionCount); 853 } 854 assert(Size > 0 && "a chain of zero size"); 855 ChainDensity[&Chain] = ExecutionCount / Size; 856 } 857 } 858 859 // Sorting chains by density in the decreasing order 860 std::stable_sort(SortedChains.begin(), SortedChains.end(), 861 [&](const Chain *C1, const Chain *C2) { 862 // Makre sure the original entry block is at the 863 // beginning of the order 864 if (C1->isEntry() != C2->isEntry()) { 865 return C1->isEntry(); 866 } 867 868 const double D1 = ChainDensity[C1]; 869 const double D2 = ChainDensity[C2]; 870 // Compare by density and break ties by chain identifiers 871 return (D1 != D2) ? (D1 > D2) : (C1->id() < C2->id()); 872 }); 873 874 // Collect the blocks in the order specified by their chains 875 Order.reserve(NumNodes); 876 for (auto Chain : SortedChains) { 877 for (auto Block : Chain->blocks()) { 878 Order.push_back(Block->Index); 879 } 880 } 881 } 882 883 private: 884 /// The number of nodes in the graph. 885 const size_t NumNodes; 886 887 /// Successors of each node. 888 std::vector<std::vector<uint64_t>> SuccNodes; 889 890 /// Predecessors of each node. 891 std::vector<std::vector<uint64_t>> PredNodes; 892 893 /// All basic blocks. 894 std::vector<Block> AllBlocks; 895 896 /// All jumps between blocks. 897 std::vector<Jump> AllJumps; 898 899 /// All chains of basic blocks. 900 std::vector<Chain> AllChains; 901 902 /// All edges between chains. 903 std::vector<ChainEdge> AllEdges; 904 905 /// Active chains. The vector gets updated at runtime when chains are merged. 906 std::vector<Chain *> HotChains; 907 }; 908 909 } // end of anonymous namespace 910 911 std::vector<uint64_t> llvm::applyExtTspLayout( 912 const std::vector<uint64_t> &NodeSizes, 913 const std::vector<uint64_t> &NodeCounts, 914 const DenseMap<std::pair<uint64_t, uint64_t>, uint64_t> &EdgeCounts) { 915 size_t NumNodes = NodeSizes.size(); 916 917 // Verify correctness of the input data. 918 assert(NodeCounts.size() == NodeSizes.size() && "Incorrect input"); 919 assert(NumNodes > 2 && "Incorrect input"); 920 921 // Apply the reordering algorithm. 922 auto Alg = ExtTSPImpl(NumNodes, NodeSizes, NodeCounts, EdgeCounts); 923 std::vector<uint64_t> Result; 924 Alg.run(Result); 925 926 // Verify correctness of the output. 927 assert(Result.front() == 0 && "Original entry point is not preserved"); 928 assert(Result.size() == NumNodes && "Incorrect size of reordered layout"); 929 return Result; 930 } 931 932 double llvm::calcExtTspScore( 933 const std::vector<uint64_t> &Order, const std::vector<uint64_t> &NodeSizes, 934 const std::vector<uint64_t> &NodeCounts, 935 const DenseMap<std::pair<uint64_t, uint64_t>, uint64_t> &EdgeCounts) { 936 // Estimate addresses of the blocks in memory 937 auto Addr = std::vector<uint64_t>(NodeSizes.size(), 0); 938 for (size_t Idx = 1; Idx < Order.size(); Idx++) { 939 Addr[Order[Idx]] = Addr[Order[Idx - 1]] + NodeSizes[Order[Idx - 1]]; 940 } 941 942 // Increase the score for each jump 943 double Score = 0; 944 for (auto It : EdgeCounts) { 945 auto Pred = It.first.first; 946 auto Succ = It.first.second; 947 uint64_t Count = It.second; 948 Score += extTSPScore(Addr[Pred], NodeSizes[Pred], Addr[Succ], Count); 949 } 950 return Score; 951 } 952 953 double llvm::calcExtTspScore( 954 const std::vector<uint64_t> &NodeSizes, 955 const std::vector<uint64_t> &NodeCounts, 956 const DenseMap<std::pair<uint64_t, uint64_t>, uint64_t> &EdgeCounts) { 957 auto Order = std::vector<uint64_t>(NodeSizes.size()); 958 for (size_t Idx = 0; Idx < NodeSizes.size(); Idx++) { 959 Order[Idx] = Idx; 960 } 961 return calcExtTspScore(Order, NodeSizes, NodeCounts, EdgeCounts); 962 } 963