1 //==- BlockFrequencyInfoImpl.h - Block Frequency Implementation -*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // Shared implementation of BlockFrequency for IR and Machine Instructions. 11 // See the documentation below for BlockFrequencyInfoImpl for details. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H 16 #define LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H 17 18 #include "llvm/ADT/DenseMap.h" 19 #include "llvm/ADT/PostOrderIterator.h" 20 #include "llvm/ADT/iterator_range.h" 21 #include "llvm/IR/BasicBlock.h" 22 #include "llvm/Support/BlockFrequency.h" 23 #include "llvm/Support/BranchProbability.h" 24 #include "llvm/Support/Debug.h" 25 #include "llvm/Support/ScaledNumber.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include <deque> 28 #include <list> 29 #include <string> 30 #include <vector> 31 32 #define DEBUG_TYPE "block-freq" 33 34 //===----------------------------------------------------------------------===// 35 // 36 // BlockMass definition. 37 // 38 // TODO: Make this private to BlockFrequencyInfoImpl or delete. 39 // 40 //===----------------------------------------------------------------------===// 41 namespace llvm { 42 43 /// \brief Mass of a block. 44 /// 45 /// This class implements a sort of fixed-point fraction always between 0.0 and 46 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0. 47 /// 48 /// Masses can be added and subtracted. Simple saturation arithmetic is used, 49 /// so arithmetic operations never overflow or underflow. 50 /// 51 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses 52 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not 53 /// quite, maximum precision). 54 /// 55 /// Masses can be scaled by \a BranchProbability at maximum precision. 56 class BlockMass { 57 uint64_t Mass; 58 59 public: 60 BlockMass() : Mass(0) {} 61 explicit BlockMass(uint64_t Mass) : Mass(Mass) {} 62 63 static BlockMass getEmpty() { return BlockMass(); } 64 static BlockMass getFull() { return BlockMass(UINT64_MAX); } 65 66 uint64_t getMass() const { return Mass; } 67 68 bool isFull() const { return Mass == UINT64_MAX; } 69 bool isEmpty() const { return !Mass; } 70 71 bool operator!() const { return isEmpty(); } 72 73 /// \brief Add another mass. 74 /// 75 /// Adds another mass, saturating at \a isFull() rather than overflowing. 76 BlockMass &operator+=(const BlockMass &X) { 77 uint64_t Sum = Mass + X.Mass; 78 Mass = Sum < Mass ? UINT64_MAX : Sum; 79 return *this; 80 } 81 82 /// \brief Subtract another mass. 83 /// 84 /// Subtracts another mass, saturating at \a isEmpty() rather than 85 /// undeflowing. 86 BlockMass &operator-=(const BlockMass &X) { 87 uint64_t Diff = Mass - X.Mass; 88 Mass = Diff > Mass ? 0 : Diff; 89 return *this; 90 } 91 92 BlockMass &operator*=(const BranchProbability &P) { 93 Mass = P.scale(Mass); 94 return *this; 95 } 96 97 bool operator==(const BlockMass &X) const { return Mass == X.Mass; } 98 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; } 99 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; } 100 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; } 101 bool operator<(const BlockMass &X) const { return Mass < X.Mass; } 102 bool operator>(const BlockMass &X) const { return Mass > X.Mass; } 103 104 /// \brief Convert to scaled number. 105 /// 106 /// Convert to \a ScaledNumber. \a isFull() gives 1.0, while \a isEmpty() 107 /// gives slightly above 0.0. 108 ScaledNumber<uint64_t> toScaled() const; 109 110 void dump() const; 111 raw_ostream &print(raw_ostream &OS) const; 112 }; 113 114 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) { 115 return BlockMass(L) += R; 116 } 117 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) { 118 return BlockMass(L) -= R; 119 } 120 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) { 121 return BlockMass(L) *= R; 122 } 123 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) { 124 return BlockMass(R) *= L; 125 } 126 127 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) { 128 return X.print(OS); 129 } 130 131 template <> struct isPodLike<BlockMass> { 132 static const bool value = true; 133 }; 134 } 135 136 //===----------------------------------------------------------------------===// 137 // 138 // BlockFrequencyInfoImpl definition. 139 // 140 //===----------------------------------------------------------------------===// 141 namespace llvm { 142 143 class BasicBlock; 144 class BranchProbabilityInfo; 145 class Function; 146 class Loop; 147 class LoopInfo; 148 class MachineBasicBlock; 149 class MachineBranchProbabilityInfo; 150 class MachineFunction; 151 class MachineLoop; 152 class MachineLoopInfo; 153 154 namespace bfi_detail { 155 struct IrreducibleGraph; 156 157 // This is part of a workaround for a GCC 4.7 crash on lambdas. 158 template <class BT> struct BlockEdgesAdder; 159 } 160 161 /// \brief Base class for BlockFrequencyInfoImpl 162 /// 163 /// BlockFrequencyInfoImplBase has supporting data structures and some 164 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on 165 /// the block type (or that call such algorithms) are skipped here. 166 /// 167 /// Nevertheless, the majority of the overall algorithm documention lives with 168 /// BlockFrequencyInfoImpl. See there for details. 169 class BlockFrequencyInfoImplBase { 170 public: 171 typedef ScaledNumber<uint64_t> Scaled64; 172 173 /// \brief Representative of a block. 174 /// 175 /// This is a simple wrapper around an index into the reverse-post-order 176 /// traversal of the blocks. 177 /// 178 /// Unlike a block pointer, its order has meaning (location in the 179 /// topological sort) and it's class is the same regardless of block type. 180 struct BlockNode { 181 typedef uint32_t IndexType; 182 IndexType Index; 183 184 bool operator==(const BlockNode &X) const { return Index == X.Index; } 185 bool operator!=(const BlockNode &X) const { return Index != X.Index; } 186 bool operator<=(const BlockNode &X) const { return Index <= X.Index; } 187 bool operator>=(const BlockNode &X) const { return Index >= X.Index; } 188 bool operator<(const BlockNode &X) const { return Index < X.Index; } 189 bool operator>(const BlockNode &X) const { return Index > X.Index; } 190 191 BlockNode() : Index(UINT32_MAX) {} 192 BlockNode(IndexType Index) : Index(Index) {} 193 194 bool isValid() const { return Index <= getMaxIndex(); } 195 static size_t getMaxIndex() { return UINT32_MAX - 1; } 196 }; 197 198 /// \brief Stats about a block itself. 199 struct FrequencyData { 200 Scaled64 Scaled; 201 uint64_t Integer; 202 }; 203 204 /// \brief Data about a loop. 205 /// 206 /// Contains the data necessary to represent represent a loop as a 207 /// pseudo-node once it's packaged. 208 struct LoopData { 209 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap; 210 typedef SmallVector<BlockNode, 4> NodeList; 211 LoopData *Parent; ///< The parent loop. 212 bool IsPackaged; ///< Whether this has been packaged. 213 uint32_t NumHeaders; ///< Number of headers. 214 ExitMap Exits; ///< Successor edges (and weights). 215 NodeList Nodes; ///< Header and the members of the loop. 216 BlockMass BackedgeMass; ///< Mass returned to loop header. 217 BlockMass Mass; 218 Scaled64 Scale; 219 220 LoopData(LoopData *Parent, const BlockNode &Header) 221 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {} 222 template <class It1, class It2> 223 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther, 224 It2 LastOther) 225 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) { 226 NumHeaders = Nodes.size(); 227 Nodes.insert(Nodes.end(), FirstOther, LastOther); 228 } 229 bool isHeader(const BlockNode &Node) const { 230 if (isIrreducible()) 231 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders, 232 Node); 233 return Node == Nodes[0]; 234 } 235 BlockNode getHeader() const { return Nodes[0]; } 236 bool isIrreducible() const { return NumHeaders > 1; } 237 238 NodeList::const_iterator members_begin() const { 239 return Nodes.begin() + NumHeaders; 240 } 241 NodeList::const_iterator members_end() const { return Nodes.end(); } 242 iterator_range<NodeList::const_iterator> members() const { 243 return make_range(members_begin(), members_end()); 244 } 245 }; 246 247 /// \brief Index of loop information. 248 struct WorkingData { 249 BlockNode Node; ///< This node. 250 LoopData *Loop; ///< The loop this block is inside. 251 BlockMass Mass; ///< Mass distribution from the entry block. 252 253 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {} 254 255 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); } 256 bool isDoubleLoopHeader() const { 257 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() && 258 Loop->Parent->isHeader(Node); 259 } 260 261 LoopData *getContainingLoop() const { 262 if (!isLoopHeader()) 263 return Loop; 264 if (!isDoubleLoopHeader()) 265 return Loop->Parent; 266 return Loop->Parent->Parent; 267 } 268 269 /// \brief Resolve a node to its representative. 270 /// 271 /// Get the node currently representing Node, which could be a containing 272 /// loop. 273 /// 274 /// This function should only be called when distributing mass. As long as 275 /// there are no irreducilbe edges to Node, then it will have complexity 276 /// O(1) in this context. 277 /// 278 /// In general, the complexity is O(L), where L is the number of loop 279 /// headers Node has been packaged into. Since this method is called in 280 /// the context of distributing mass, L will be the number of loop headers 281 /// an early exit edge jumps out of. 282 BlockNode getResolvedNode() const { 283 auto L = getPackagedLoop(); 284 return L ? L->getHeader() : Node; 285 } 286 LoopData *getPackagedLoop() const { 287 if (!Loop || !Loop->IsPackaged) 288 return nullptr; 289 auto L = Loop; 290 while (L->Parent && L->Parent->IsPackaged) 291 L = L->Parent; 292 return L; 293 } 294 295 /// \brief Get the appropriate mass for a node. 296 /// 297 /// Get appropriate mass for Node. If Node is a loop-header (whose loop 298 /// has been packaged), returns the mass of its pseudo-node. If it's a 299 /// node inside a packaged loop, it returns the loop's mass. 300 BlockMass &getMass() { 301 if (!isAPackage()) 302 return Mass; 303 if (!isADoublePackage()) 304 return Loop->Mass; 305 return Loop->Parent->Mass; 306 } 307 308 /// \brief Has ContainingLoop been packaged up? 309 bool isPackaged() const { return getResolvedNode() != Node; } 310 /// \brief Has Loop been packaged up? 311 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; } 312 /// \brief Has Loop been packaged up twice? 313 bool isADoublePackage() const { 314 return isDoubleLoopHeader() && Loop->Parent->IsPackaged; 315 } 316 }; 317 318 /// \brief Unscaled probability weight. 319 /// 320 /// Probability weight for an edge in the graph (including the 321 /// successor/target node). 322 /// 323 /// All edges in the original function are 32-bit. However, exit edges from 324 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of 325 /// space in general. 326 /// 327 /// In addition to the raw weight amount, Weight stores the type of the edge 328 /// in the current context (i.e., the context of the loop being processed). 329 /// Is this a local edge within the loop, an exit from the loop, or a 330 /// backedge to the loop header? 331 struct Weight { 332 enum DistType { Local, Exit, Backedge }; 333 DistType Type; 334 BlockNode TargetNode; 335 uint64_t Amount; 336 Weight() : Type(Local), Amount(0) {} 337 }; 338 339 /// \brief Distribution of unscaled probability weight. 340 /// 341 /// Distribution of unscaled probability weight to a set of successors. 342 /// 343 /// This class collates the successor edge weights for later processing. 344 /// 345 /// \a DidOverflow indicates whether \a Total did overflow while adding to 346 /// the distribution. It should never overflow twice. 347 struct Distribution { 348 typedef SmallVector<Weight, 4> WeightList; 349 WeightList Weights; ///< Individual successor weights. 350 uint64_t Total; ///< Sum of all weights. 351 bool DidOverflow; ///< Whether \a Total did overflow. 352 353 Distribution() : Total(0), DidOverflow(false) {} 354 void addLocal(const BlockNode &Node, uint64_t Amount) { 355 add(Node, Amount, Weight::Local); 356 } 357 void addExit(const BlockNode &Node, uint64_t Amount) { 358 add(Node, Amount, Weight::Exit); 359 } 360 void addBackedge(const BlockNode &Node, uint64_t Amount) { 361 add(Node, Amount, Weight::Backedge); 362 } 363 364 /// \brief Normalize the distribution. 365 /// 366 /// Combines multiple edges to the same \a Weight::TargetNode and scales 367 /// down so that \a Total fits into 32-bits. 368 /// 369 /// This is linear in the size of \a Weights. For the vast majority of 370 /// cases, adjacent edge weights are combined by sorting WeightList and 371 /// combining adjacent weights. However, for very large edge lists an 372 /// auxiliary hash table is used. 373 void normalize(); 374 375 private: 376 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type); 377 }; 378 379 /// \brief Data about each block. This is used downstream. 380 std::vector<FrequencyData> Freqs; 381 382 /// \brief Loop data: see initializeLoops(). 383 std::vector<WorkingData> Working; 384 385 /// \brief Indexed information about loops. 386 std::list<LoopData> Loops; 387 388 /// \brief Add all edges out of a packaged loop to the distribution. 389 /// 390 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each 391 /// successor edge. 392 /// 393 /// \return \c true unless there's an irreducible backedge. 394 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop, 395 Distribution &Dist); 396 397 /// \brief Add an edge to the distribution. 398 /// 399 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the 400 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise, 401 /// every edge should be a local edge (since all the loops are packaged up). 402 /// 403 /// \return \c true unless aborted due to an irreducible backedge. 404 bool addToDist(Distribution &Dist, const LoopData *OuterLoop, 405 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight); 406 407 LoopData &getLoopPackage(const BlockNode &Head) { 408 assert(Head.Index < Working.size()); 409 assert(Working[Head.Index].isLoopHeader()); 410 return *Working[Head.Index].Loop; 411 } 412 413 /// \brief Analyze irreducible SCCs. 414 /// 415 /// Separate irreducible SCCs from \c G, which is an explict graph of \c 416 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr). 417 /// Insert them into \a Loops before \c Insert. 418 /// 419 /// \return the \c LoopData nodes representing the irreducible SCCs. 420 iterator_range<std::list<LoopData>::iterator> 421 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop, 422 std::list<LoopData>::iterator Insert); 423 424 /// \brief Update a loop after packaging irreducible SCCs inside of it. 425 /// 426 /// Update \c OuterLoop. Before finding irreducible control flow, it was 427 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a 428 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged 429 /// up need to be removed from \a OuterLoop::Nodes. 430 void updateLoopWithIrreducible(LoopData &OuterLoop); 431 432 /// \brief Distribute mass according to a distribution. 433 /// 434 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(), 435 /// backedges and exits are stored in its entry in Loops. 436 /// 437 /// Mass is distributed in parallel from two copies of the source mass. 438 void distributeMass(const BlockNode &Source, LoopData *OuterLoop, 439 Distribution &Dist); 440 441 /// \brief Compute the loop scale for a loop. 442 void computeLoopScale(LoopData &Loop); 443 444 /// \brief Package up a loop. 445 void packageLoop(LoopData &Loop); 446 447 /// \brief Unwrap loops. 448 void unwrapLoops(); 449 450 /// \brief Finalize frequency metrics. 451 /// 452 /// Calculates final frequencies and cleans up no-longer-needed data 453 /// structures. 454 void finalizeMetrics(); 455 456 /// \brief Clear all memory. 457 void clear(); 458 459 virtual std::string getBlockName(const BlockNode &Node) const; 460 std::string getLoopName(const LoopData &Loop) const; 461 462 virtual raw_ostream &print(raw_ostream &OS) const { return OS; } 463 void dump() const { print(dbgs()); } 464 465 Scaled64 getFloatingBlockFreq(const BlockNode &Node) const; 466 467 BlockFrequency getBlockFreq(const BlockNode &Node) const; 468 469 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const; 470 raw_ostream &printBlockFreq(raw_ostream &OS, 471 const BlockFrequency &Freq) const; 472 473 uint64_t getEntryFreq() const { 474 assert(!Freqs.empty()); 475 return Freqs[0].Integer; 476 } 477 /// \brief Virtual destructor. 478 /// 479 /// Need a virtual destructor to mask the compiler warning about 480 /// getBlockName(). 481 virtual ~BlockFrequencyInfoImplBase() {} 482 }; 483 484 namespace bfi_detail { 485 template <class BlockT> struct TypeMap {}; 486 template <> struct TypeMap<BasicBlock> { 487 typedef BasicBlock BlockT; 488 typedef Function FunctionT; 489 typedef BranchProbabilityInfo BranchProbabilityInfoT; 490 typedef Loop LoopT; 491 typedef LoopInfo LoopInfoT; 492 }; 493 template <> struct TypeMap<MachineBasicBlock> { 494 typedef MachineBasicBlock BlockT; 495 typedef MachineFunction FunctionT; 496 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT; 497 typedef MachineLoop LoopT; 498 typedef MachineLoopInfo LoopInfoT; 499 }; 500 501 /// \brief Get the name of a MachineBasicBlock. 502 /// 503 /// Get the name of a MachineBasicBlock. It's templated so that including from 504 /// CodeGen is unnecessary (that would be a layering issue). 505 /// 506 /// This is used mainly for debug output. The name is similar to 507 /// MachineBasicBlock::getFullName(), but skips the name of the function. 508 template <class BlockT> std::string getBlockName(const BlockT *BB) { 509 assert(BB && "Unexpected nullptr"); 510 auto MachineName = "BB" + Twine(BB->getNumber()); 511 if (BB->getBasicBlock()) 512 return (MachineName + "[" + BB->getName() + "]").str(); 513 return MachineName.str(); 514 } 515 /// \brief Get the name of a BasicBlock. 516 template <> inline std::string getBlockName(const BasicBlock *BB) { 517 assert(BB && "Unexpected nullptr"); 518 return BB->getName().str(); 519 } 520 521 /// \brief Graph of irreducible control flow. 522 /// 523 /// This graph is used for determining the SCCs in a loop (or top-level 524 /// function) that has irreducible control flow. 525 /// 526 /// During the block frequency algorithm, the local graphs are defined in a 527 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock 528 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The 529 /// latter only has successor information. 530 /// 531 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use 532 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator), 533 /// and it explicitly lists predecessors and successors. The initialization 534 /// that relies on \c MachineBasicBlock is defined in the header. 535 struct IrreducibleGraph { 536 typedef BlockFrequencyInfoImplBase BFIBase; 537 538 BFIBase &BFI; 539 540 typedef BFIBase::BlockNode BlockNode; 541 struct IrrNode { 542 BlockNode Node; 543 unsigned NumIn; 544 std::deque<const IrrNode *> Edges; 545 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {} 546 547 typedef std::deque<const IrrNode *>::const_iterator iterator; 548 iterator pred_begin() const { return Edges.begin(); } 549 iterator succ_begin() const { return Edges.begin() + NumIn; } 550 iterator pred_end() const { return succ_begin(); } 551 iterator succ_end() const { return Edges.end(); } 552 }; 553 BlockNode Start; 554 const IrrNode *StartIrr; 555 std::vector<IrrNode> Nodes; 556 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup; 557 558 /// \brief Construct an explicit graph containing irreducible control flow. 559 /// 560 /// Construct an explicit graph of the control flow in \c OuterLoop (or the 561 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c 562 /// addBlockEdges to add block successors that have not been packaged into 563 /// loops. 564 /// 565 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected 566 /// user of this. 567 template <class BlockEdgesAdder> 568 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop, 569 BlockEdgesAdder addBlockEdges) 570 : BFI(BFI), StartIrr(nullptr) { 571 initialize(OuterLoop, addBlockEdges); 572 } 573 574 template <class BlockEdgesAdder> 575 void initialize(const BFIBase::LoopData *OuterLoop, 576 BlockEdgesAdder addBlockEdges); 577 void addNodesInLoop(const BFIBase::LoopData &OuterLoop); 578 void addNodesInFunction(); 579 void addNode(const BlockNode &Node) { 580 Nodes.emplace_back(Node); 581 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty(); 582 } 583 void indexNodes(); 584 template <class BlockEdgesAdder> 585 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop, 586 BlockEdgesAdder addBlockEdges); 587 void addEdge(IrrNode &Irr, const BlockNode &Succ, 588 const BFIBase::LoopData *OuterLoop); 589 }; 590 template <class BlockEdgesAdder> 591 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop, 592 BlockEdgesAdder addBlockEdges) { 593 if (OuterLoop) { 594 addNodesInLoop(*OuterLoop); 595 for (auto N : OuterLoop->Nodes) 596 addEdges(N, OuterLoop, addBlockEdges); 597 } else { 598 addNodesInFunction(); 599 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index) 600 addEdges(Index, OuterLoop, addBlockEdges); 601 } 602 StartIrr = Lookup[Start.Index]; 603 } 604 template <class BlockEdgesAdder> 605 void IrreducibleGraph::addEdges(const BlockNode &Node, 606 const BFIBase::LoopData *OuterLoop, 607 BlockEdgesAdder addBlockEdges) { 608 auto L = Lookup.find(Node.Index); 609 if (L == Lookup.end()) 610 return; 611 IrrNode &Irr = *L->second; 612 const auto &Working = BFI.Working[Node.Index]; 613 614 if (Working.isAPackage()) 615 for (const auto &I : Working.Loop->Exits) 616 addEdge(Irr, I.first, OuterLoop); 617 else 618 addBlockEdges(*this, Irr, OuterLoop); 619 } 620 } 621 622 /// \brief Shared implementation for block frequency analysis. 623 /// 624 /// This is a shared implementation of BlockFrequencyInfo and 625 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of 626 /// blocks. 627 /// 628 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block, 629 /// which is called the header. A given loop, L, can have sub-loops, which are 630 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC 631 /// consists of a single block that does not have a self-edge.) 632 /// 633 /// In addition to loops, this algorithm has limited support for irreducible 634 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are 635 /// discovered on they fly, and modelled as loops with multiple headers. 636 /// 637 /// The headers of irreducible sub-SCCs consist of its entry blocks and all 638 /// nodes that are targets of a backedge within it (excluding backedges within 639 /// true sub-loops). Block frequency calculations act as if a block is 640 /// inserted that intercepts all the edges to the headers. All backedges and 641 /// entries point to this block. Its successors are the headers, which split 642 /// the frequency evenly. 643 /// 644 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision, 645 /// separates mass distribution from loop scaling, and dithers to eliminate 646 /// probability mass loss. 647 /// 648 /// The implementation is split between BlockFrequencyInfoImpl, which knows the 649 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and 650 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a 651 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in 652 /// reverse-post order. This gives two advantages: it's easy to compare the 653 /// relative ordering of two nodes, and maps keyed on BlockT can be represented 654 /// by vectors. 655 /// 656 /// This algorithm is O(V+E), unless there is irreducible control flow, in 657 /// which case it's O(V*E) in the worst case. 658 /// 659 /// These are the main stages: 660 /// 661 /// 0. Reverse post-order traversal (\a initializeRPOT()). 662 /// 663 /// Run a single post-order traversal and save it (in reverse) in RPOT. 664 /// All other stages make use of this ordering. Save a lookup from BlockT 665 /// to BlockNode (the index into RPOT) in Nodes. 666 /// 667 /// 1. Loop initialization (\a initializeLoops()). 668 /// 669 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of 670 /// the algorithm. In particular, store the immediate members of each loop 671 /// in reverse post-order. 672 /// 673 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()). 674 /// 675 /// For each loop (bottom-up), distribute mass through the DAG resulting 676 /// from ignoring backedges and treating sub-loops as a single pseudo-node. 677 /// Track the backedge mass distributed to the loop header, and use it to 678 /// calculate the loop scale (number of loop iterations). Immediate 679 /// members that represent sub-loops will already have been visited and 680 /// packaged into a pseudo-node. 681 /// 682 /// Distributing mass in a loop is a reverse-post-order traversal through 683 /// the loop. Start by assigning full mass to the Loop header. For each 684 /// node in the loop: 685 /// 686 /// - Fetch and categorize the weight distribution for its successors. 687 /// If this is a packaged-subloop, the weight distribution is stored 688 /// in \a LoopData::Exits. Otherwise, fetch it from 689 /// BranchProbabilityInfo. 690 /// 691 /// - Each successor is categorized as \a Weight::Local, a local edge 692 /// within the current loop, \a Weight::Backedge, a backedge to the 693 /// loop header, or \a Weight::Exit, any successor outside the loop. 694 /// The weight, the successor, and its category are stored in \a 695 /// Distribution. There can be multiple edges to each successor. 696 /// 697 /// - If there's a backedge to a non-header, there's an irreducible SCC. 698 /// The usual flow is temporarily aborted. \a 699 /// computeIrreducibleMass() finds the irreducible SCCs within the 700 /// loop, packages them up, and restarts the flow. 701 /// 702 /// - Normalize the distribution: scale weights down so that their sum 703 /// is 32-bits, and coalesce multiple edges to the same node. 704 /// 705 /// - Distribute the mass accordingly, dithering to minimize mass loss, 706 /// as described in \a distributeMass(). 707 /// 708 /// Finally, calculate the loop scale from the accumulated backedge mass. 709 /// 710 /// 3. Distribute mass in the function (\a computeMassInFunction()). 711 /// 712 /// Finally, distribute mass through the DAG resulting from packaging all 713 /// loops in the function. This uses the same algorithm as distributing 714 /// mass in a loop, except that there are no exit or backedge edges. 715 /// 716 /// 4. Unpackage loops (\a unwrapLoops()). 717 /// 718 /// Initialize each block's frequency to a floating point representation of 719 /// its mass. 720 /// 721 /// Visit loops top-down, scaling the frequencies of its immediate members 722 /// by the loop's pseudo-node's frequency. 723 /// 724 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()). 725 /// 726 /// Using the min and max frequencies as a guide, translate floating point 727 /// frequencies to an appropriate range in uint64_t. 728 /// 729 /// It has some known flaws. 730 /// 731 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting 732 /// BlockFrequency's 64-bit integer precision. 733 /// 734 /// - The model of irreducible control flow is a rough approximation. 735 /// 736 /// Modelling irreducible control flow exactly involves setting up and 737 /// solving a group of infinite geometric series. Such precision is 738 /// unlikely to be worthwhile, since most of our algorithms give up on 739 /// irreducible control flow anyway. 740 /// 741 /// Nevertheless, we might find that we need to get closer. Here's a sort 742 /// of TODO list for the model with diminishing returns, to be completed as 743 /// necessary. 744 /// 745 /// - The headers for the \a LoopData representing an irreducible SCC 746 /// include non-entry blocks. When these extra blocks exist, they 747 /// indicate a self-contained irreducible sub-SCC. We could treat them 748 /// as sub-loops, rather than arbitrarily shoving the problematic 749 /// blocks into the headers of the main irreducible SCC. 750 /// 751 /// - Backedge frequencies are assumed to be evenly split between the 752 /// headers of a given irreducible SCC. Instead, we could track the 753 /// backedge mass separately for each header, and adjust their relative 754 /// frequencies. 755 /// 756 /// - Entry frequencies are assumed to be evenly split between the 757 /// headers of a given irreducible SCC, which is the only option if we 758 /// need to compute mass in the SCC before its parent loop. Instead, 759 /// we could partially compute mass in the parent loop, and stop when 760 /// we get to the SCC. Here, we have the correct ratio of entry 761 /// masses, which we can use to adjust their relative frequencies. 762 /// Compute mass in the SCC, and then continue propagation in the 763 /// parent. 764 /// 765 /// - We can propagate mass iteratively through the SCC, for some fixed 766 /// number of iterations. Each iteration starts by assigning the entry 767 /// blocks their backedge mass from the prior iteration. The final 768 /// mass for each block (and each exit, and the total backedge mass 769 /// used for computing loop scale) is the sum of all iterations. 770 /// (Running this until fixed point would "solve" the geometric 771 /// series by simulation.) 772 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase { 773 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT; 774 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT; 775 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT 776 BranchProbabilityInfoT; 777 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT; 778 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT; 779 780 // This is part of a workaround for a GCC 4.7 crash on lambdas. 781 friend struct bfi_detail::BlockEdgesAdder<BT>; 782 783 typedef GraphTraits<const BlockT *> Successor; 784 typedef GraphTraits<Inverse<const BlockT *>> Predecessor; 785 786 const BranchProbabilityInfoT *BPI; 787 const LoopInfoT *LI; 788 const FunctionT *F; 789 790 // All blocks in reverse postorder. 791 std::vector<const BlockT *> RPOT; 792 DenseMap<const BlockT *, BlockNode> Nodes; 793 794 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator; 795 796 rpot_iterator rpot_begin() const { return RPOT.begin(); } 797 rpot_iterator rpot_end() const { return RPOT.end(); } 798 799 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); } 800 801 BlockNode getNode(const rpot_iterator &I) const { 802 return BlockNode(getIndex(I)); 803 } 804 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); } 805 806 const BlockT *getBlock(const BlockNode &Node) const { 807 assert(Node.Index < RPOT.size()); 808 return RPOT[Node.Index]; 809 } 810 811 /// \brief Run (and save) a post-order traversal. 812 /// 813 /// Saves a reverse post-order traversal of all the nodes in \a F. 814 void initializeRPOT(); 815 816 /// \brief Initialize loop data. 817 /// 818 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from 819 /// each block to the deepest loop it's in, but we need the inverse. For each 820 /// loop, we store in reverse post-order its "immediate" members, defined as 821 /// the header, the headers of immediate sub-loops, and all other blocks in 822 /// the loop that are not in sub-loops. 823 void initializeLoops(); 824 825 /// \brief Propagate to a block's successors. 826 /// 827 /// In the context of distributing mass through \c OuterLoop, divide the mass 828 /// currently assigned to \c Node between its successors. 829 /// 830 /// \return \c true unless there's an irreducible backedge. 831 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node); 832 833 /// \brief Compute mass in a particular loop. 834 /// 835 /// Assign mass to \c Loop's header, and then for each block in \c Loop in 836 /// reverse post-order, distribute mass to its successors. Only visits nodes 837 /// that have not been packaged into sub-loops. 838 /// 839 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop. 840 /// \return \c true unless there's an irreducible backedge. 841 bool computeMassInLoop(LoopData &Loop); 842 843 /// \brief Try to compute mass in the top-level function. 844 /// 845 /// Assign mass to the entry block, and then for each block in reverse 846 /// post-order, distribute mass to its successors. Skips nodes that have 847 /// been packaged into loops. 848 /// 849 /// \pre \a computeMassInLoops() has been called. 850 /// \return \c true unless there's an irreducible backedge. 851 bool tryToComputeMassInFunction(); 852 853 /// \brief Compute mass in (and package up) irreducible SCCs. 854 /// 855 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front 856 /// of \c Insert), and call \a computeMassInLoop() on each of them. 857 /// 858 /// If \c OuterLoop is \c nullptr, it refers to the top-level function. 859 /// 860 /// \pre \a computeMassInLoop() has been called for each subloop of \c 861 /// OuterLoop. 862 /// \pre \c Insert points at the the last loop successfully processed by \a 863 /// computeMassInLoop(). 864 /// \pre \c OuterLoop has irreducible SCCs. 865 void computeIrreducibleMass(LoopData *OuterLoop, 866 std::list<LoopData>::iterator Insert); 867 868 /// \brief Compute mass in all loops. 869 /// 870 /// For each loop bottom-up, call \a computeMassInLoop(). 871 /// 872 /// \a computeMassInLoop() aborts (and returns \c false) on loops that 873 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then 874 /// re-enter \a computeMassInLoop(). 875 /// 876 /// \post \a computeMassInLoop() has returned \c true for every loop. 877 void computeMassInLoops(); 878 879 /// \brief Compute mass in the top-level function. 880 /// 881 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to 882 /// compute mass in the top-level function. 883 /// 884 /// \post \a tryToComputeMassInFunction() has returned \c true. 885 void computeMassInFunction(); 886 887 std::string getBlockName(const BlockNode &Node) const override { 888 return bfi_detail::getBlockName(getBlock(Node)); 889 } 890 891 public: 892 const FunctionT *getFunction() const { return F; } 893 894 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI, 895 const LoopInfoT *LI); 896 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {} 897 898 using BlockFrequencyInfoImplBase::getEntryFreq; 899 BlockFrequency getBlockFreq(const BlockT *BB) const { 900 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB)); 901 } 902 Scaled64 getFloatingBlockFreq(const BlockT *BB) const { 903 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB)); 904 } 905 906 /// \brief Print the frequencies for the current function. 907 /// 908 /// Prints the frequencies for the blocks in the current function. 909 /// 910 /// Blocks are printed in the natural iteration order of the function, rather 911 /// than reverse post-order. This provides two advantages: writing -analyze 912 /// tests is easier (since blocks come out in source order), and even 913 /// unreachable blocks are printed. 914 /// 915 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so 916 /// we need to override it here. 917 raw_ostream &print(raw_ostream &OS) const override; 918 using BlockFrequencyInfoImplBase::dump; 919 920 using BlockFrequencyInfoImplBase::printBlockFreq; 921 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const { 922 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB)); 923 } 924 }; 925 926 template <class BT> 927 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F, 928 const BranchProbabilityInfoT *BPI, 929 const LoopInfoT *LI) { 930 // Save the parameters. 931 this->BPI = BPI; 932 this->LI = LI; 933 this->F = F; 934 935 // Clean up left-over data structures. 936 BlockFrequencyInfoImplBase::clear(); 937 RPOT.clear(); 938 Nodes.clear(); 939 940 // Initialize. 941 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n=================" 942 << std::string(F->getName().size(), '=') << "\n"); 943 initializeRPOT(); 944 initializeLoops(); 945 946 // Visit loops in post-order to find thelocal mass distribution, and then do 947 // the full function. 948 computeMassInLoops(); 949 computeMassInFunction(); 950 unwrapLoops(); 951 finalizeMetrics(); 952 } 953 954 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() { 955 const BlockT *Entry = F->begin(); 956 RPOT.reserve(F->size()); 957 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT)); 958 std::reverse(RPOT.begin(), RPOT.end()); 959 960 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() && 961 "More nodes in function than Block Frequency Info supports"); 962 963 DEBUG(dbgs() << "reverse-post-order-traversal\n"); 964 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) { 965 BlockNode Node = getNode(I); 966 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n"); 967 Nodes[*I] = Node; 968 } 969 970 Working.reserve(RPOT.size()); 971 for (size_t Index = 0; Index < RPOT.size(); ++Index) 972 Working.emplace_back(Index); 973 Freqs.resize(RPOT.size()); 974 } 975 976 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() { 977 DEBUG(dbgs() << "loop-detection\n"); 978 if (LI->empty()) 979 return; 980 981 // Visit loops top down and assign them an index. 982 std::deque<std::pair<const LoopT *, LoopData *>> Q; 983 for (const LoopT *L : *LI) 984 Q.emplace_back(L, nullptr); 985 while (!Q.empty()) { 986 const LoopT *Loop = Q.front().first; 987 LoopData *Parent = Q.front().second; 988 Q.pop_front(); 989 990 BlockNode Header = getNode(Loop->getHeader()); 991 assert(Header.isValid()); 992 993 Loops.emplace_back(Parent, Header); 994 Working[Header.Index].Loop = &Loops.back(); 995 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n"); 996 997 for (const LoopT *L : *Loop) 998 Q.emplace_back(L, &Loops.back()); 999 } 1000 1001 // Visit nodes in reverse post-order and add them to their deepest containing 1002 // loop. 1003 for (size_t Index = 0; Index < RPOT.size(); ++Index) { 1004 // Loop headers have already been mostly mapped. 1005 if (Working[Index].isLoopHeader()) { 1006 LoopData *ContainingLoop = Working[Index].getContainingLoop(); 1007 if (ContainingLoop) 1008 ContainingLoop->Nodes.push_back(Index); 1009 continue; 1010 } 1011 1012 const LoopT *Loop = LI->getLoopFor(RPOT[Index]); 1013 if (!Loop) 1014 continue; 1015 1016 // Add this node to its containing loop's member list. 1017 BlockNode Header = getNode(Loop->getHeader()); 1018 assert(Header.isValid()); 1019 const auto &HeaderData = Working[Header.Index]; 1020 assert(HeaderData.isLoopHeader()); 1021 1022 Working[Index].Loop = HeaderData.Loop; 1023 HeaderData.Loop->Nodes.push_back(Index); 1024 DEBUG(dbgs() << " - loop = " << getBlockName(Header) 1025 << ": member = " << getBlockName(Index) << "\n"); 1026 } 1027 } 1028 1029 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() { 1030 // Visit loops with the deepest first, and the top-level loops last. 1031 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) { 1032 if (computeMassInLoop(*L)) 1033 continue; 1034 auto Next = std::next(L); 1035 computeIrreducibleMass(&*L, L.base()); 1036 L = std::prev(Next); 1037 if (computeMassInLoop(*L)) 1038 continue; 1039 llvm_unreachable("unhandled irreducible control flow"); 1040 } 1041 } 1042 1043 template <class BT> 1044 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) { 1045 // Compute mass in loop. 1046 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n"); 1047 1048 if (Loop.isIrreducible()) { 1049 BlockMass Remaining = BlockMass::getFull(); 1050 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) { 1051 auto &Mass = Working[Loop.Nodes[H].Index].getMass(); 1052 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H); 1053 Remaining -= Mass; 1054 } 1055 for (const BlockNode &M : Loop.Nodes) 1056 if (!propagateMassToSuccessors(&Loop, M)) 1057 llvm_unreachable("unhandled irreducible control flow"); 1058 } else { 1059 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull(); 1060 if (!propagateMassToSuccessors(&Loop, Loop.getHeader())) 1061 llvm_unreachable("irreducible control flow to loop header!?"); 1062 for (const BlockNode &M : Loop.members()) 1063 if (!propagateMassToSuccessors(&Loop, M)) 1064 // Irreducible backedge. 1065 return false; 1066 } 1067 1068 computeLoopScale(Loop); 1069 packageLoop(Loop); 1070 return true; 1071 } 1072 1073 template <class BT> 1074 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() { 1075 // Compute mass in function. 1076 DEBUG(dbgs() << "compute-mass-in-function\n"); 1077 assert(!Working.empty() && "no blocks in function"); 1078 assert(!Working[0].isLoopHeader() && "entry block is a loop header"); 1079 1080 Working[0].getMass() = BlockMass::getFull(); 1081 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) { 1082 // Check for nodes that have been packaged. 1083 BlockNode Node = getNode(I); 1084 if (Working[Node.Index].isPackaged()) 1085 continue; 1086 1087 if (!propagateMassToSuccessors(nullptr, Node)) 1088 return false; 1089 } 1090 return true; 1091 } 1092 1093 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() { 1094 if (tryToComputeMassInFunction()) 1095 return; 1096 computeIrreducibleMass(nullptr, Loops.begin()); 1097 if (tryToComputeMassInFunction()) 1098 return; 1099 llvm_unreachable("unhandled irreducible control flow"); 1100 } 1101 1102 /// \note This should be a lambda, but that crashes GCC 4.7. 1103 namespace bfi_detail { 1104 template <class BT> struct BlockEdgesAdder { 1105 typedef BT BlockT; 1106 typedef BlockFrequencyInfoImplBase::LoopData LoopData; 1107 typedef GraphTraits<const BlockT *> Successor; 1108 1109 const BlockFrequencyInfoImpl<BT> &BFI; 1110 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI) 1111 : BFI(BFI) {} 1112 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr, 1113 const LoopData *OuterLoop) { 1114 const BlockT *BB = BFI.RPOT[Irr.Node.Index]; 1115 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB); 1116 I != E; ++I) 1117 G.addEdge(Irr, BFI.getNode(*I), OuterLoop); 1118 } 1119 }; 1120 } 1121 template <class BT> 1122 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass( 1123 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) { 1124 DEBUG(dbgs() << "analyze-irreducible-in-"; 1125 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n"; 1126 else dbgs() << "function\n"); 1127 1128 using namespace bfi_detail; 1129 // Ideally, addBlockEdges() would be declared here as a lambda, but that 1130 // crashes GCC 4.7. 1131 BlockEdgesAdder<BT> addBlockEdges(*this); 1132 IrreducibleGraph G(*this, OuterLoop, addBlockEdges); 1133 1134 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert)) 1135 computeMassInLoop(L); 1136 1137 if (!OuterLoop) 1138 return; 1139 updateLoopWithIrreducible(*OuterLoop); 1140 } 1141 1142 template <class BT> 1143 bool 1144 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop, 1145 const BlockNode &Node) { 1146 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n"); 1147 // Calculate probability for successors. 1148 Distribution Dist; 1149 if (auto *Loop = Working[Node.Index].getPackagedLoop()) { 1150 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop"); 1151 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist)) 1152 // Irreducible backedge. 1153 return false; 1154 } else { 1155 const BlockT *BB = getBlock(Node); 1156 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB); 1157 SI != SE; ++SI) 1158 // Do not dereference SI, or getEdgeWeight() is linear in the number of 1159 // successors. 1160 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI), 1161 BPI->getEdgeWeight(BB, SI))) 1162 // Irreducible backedge. 1163 return false; 1164 } 1165 1166 // Distribute mass to successors, saving exit and backedge data in the 1167 // loop header. 1168 distributeMass(Node, OuterLoop, Dist); 1169 return true; 1170 } 1171 1172 template <class BT> 1173 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const { 1174 if (!F) 1175 return OS; 1176 OS << "block-frequency-info: " << F->getName() << "\n"; 1177 for (const BlockT &BB : *F) 1178 OS << " - " << bfi_detail::getBlockName(&BB) 1179 << ": float = " << getFloatingBlockFreq(&BB) 1180 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n"; 1181 1182 // Add an extra newline for readability. 1183 OS << "\n"; 1184 return OS; 1185 } 1186 } 1187 1188 #undef DEBUG_TYPE 1189 1190 #endif 1191