1 //===- MachineBlockPlacement.cpp - Basic Block Code Layout optimization ---===// 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 // This file implements basic block placement transformations using the CFG 10 // structure and branch probability estimates. 11 // 12 // The pass strives to preserve the structure of the CFG (that is, retain 13 // a topological ordering of basic blocks) in the absence of a *strong* signal 14 // to the contrary from probabilities. However, within the CFG structure, it 15 // attempts to choose an ordering which favors placing more likely sequences of 16 // blocks adjacent to each other. 17 // 18 // The algorithm works from the inner-most loop within a function outward, and 19 // at each stage walks through the basic blocks, trying to coalesce them into 20 // sequential chains where allowed by the CFG (or demanded by heavy 21 // probabilities). Finally, it walks the blocks in topological order, and the 22 // first time it reaches a chain of basic blocks, it schedules them in the 23 // function in-order. 24 // 25 //===----------------------------------------------------------------------===// 26 27 #include "BranchFolding.h" 28 #include "llvm/ADT/ArrayRef.h" 29 #include "llvm/ADT/DenseMap.h" 30 #include "llvm/ADT/STLExtras.h" 31 #include "llvm/ADT/SetVector.h" 32 #include "llvm/ADT/SmallPtrSet.h" 33 #include "llvm/ADT/SmallVector.h" 34 #include "llvm/ADT/Statistic.h" 35 #include "llvm/Analysis/BlockFrequencyInfoImpl.h" 36 #include "llvm/Analysis/ProfileSummaryInfo.h" 37 #include "llvm/CodeGen/MBFIWrapper.h" 38 #include "llvm/CodeGen/MachineBasicBlock.h" 39 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 40 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 41 #include "llvm/CodeGen/MachineFunction.h" 42 #include "llvm/CodeGen/MachineFunctionPass.h" 43 #include "llvm/CodeGen/MachineLoopInfo.h" 44 #include "llvm/CodeGen/MachinePostDominators.h" 45 #include "llvm/CodeGen/MachineSizeOpts.h" 46 #include "llvm/CodeGen/TailDuplicator.h" 47 #include "llvm/CodeGen/TargetInstrInfo.h" 48 #include "llvm/CodeGen/TargetLowering.h" 49 #include "llvm/CodeGen/TargetPassConfig.h" 50 #include "llvm/CodeGen/TargetSubtargetInfo.h" 51 #include "llvm/IR/DebugLoc.h" 52 #include "llvm/IR/Function.h" 53 #include "llvm/IR/PrintPasses.h" 54 #include "llvm/InitializePasses.h" 55 #include "llvm/Pass.h" 56 #include "llvm/Support/Allocator.h" 57 #include "llvm/Support/BlockFrequency.h" 58 #include "llvm/Support/BranchProbability.h" 59 #include "llvm/Support/CodeGen.h" 60 #include "llvm/Support/CommandLine.h" 61 #include "llvm/Support/Compiler.h" 62 #include "llvm/Support/Debug.h" 63 #include "llvm/Support/raw_ostream.h" 64 #include "llvm/Target/TargetMachine.h" 65 #include "llvm/Transforms/Utils/CodeLayout.h" 66 #include <algorithm> 67 #include <cassert> 68 #include <cstdint> 69 #include <iterator> 70 #include <memory> 71 #include <string> 72 #include <tuple> 73 #include <utility> 74 #include <vector> 75 76 using namespace llvm; 77 78 #define DEBUG_TYPE "block-placement" 79 80 STATISTIC(NumCondBranches, "Number of conditional branches"); 81 STATISTIC(NumUncondBranches, "Number of unconditional branches"); 82 STATISTIC(CondBranchTakenFreq, 83 "Potential frequency of taking conditional branches"); 84 STATISTIC(UncondBranchTakenFreq, 85 "Potential frequency of taking unconditional branches"); 86 87 static cl::opt<unsigned> AlignAllBlock( 88 "align-all-blocks", 89 cl::desc("Force the alignment of all blocks in the function in log2 format " 90 "(e.g 4 means align on 16B boundaries)."), 91 cl::init(0), cl::Hidden); 92 93 static cl::opt<unsigned> AlignAllNonFallThruBlocks( 94 "align-all-nofallthru-blocks", 95 cl::desc("Force the alignment of all blocks that have no fall-through " 96 "predecessors (i.e. don't add nops that are executed). In log2 " 97 "format (e.g 4 means align on 16B boundaries)."), 98 cl::init(0), cl::Hidden); 99 100 static cl::opt<unsigned> MaxBytesForAlignmentOverride( 101 "max-bytes-for-alignment", 102 cl::desc("Forces the maximum bytes allowed to be emitted when padding for " 103 "alignment"), 104 cl::init(0), cl::Hidden); 105 106 // FIXME: Find a good default for this flag and remove the flag. 107 static cl::opt<unsigned> ExitBlockBias( 108 "block-placement-exit-block-bias", 109 cl::desc("Block frequency percentage a loop exit block needs " 110 "over the original exit to be considered the new exit."), 111 cl::init(0), cl::Hidden); 112 113 // Definition: 114 // - Outlining: placement of a basic block outside the chain or hot path. 115 116 static cl::opt<unsigned> LoopToColdBlockRatio( 117 "loop-to-cold-block-ratio", 118 cl::desc("Outline loop blocks from loop chain if (frequency of loop) / " 119 "(frequency of block) is greater than this ratio"), 120 cl::init(5), cl::Hidden); 121 122 static cl::opt<bool> ForceLoopColdBlock( 123 "force-loop-cold-block", 124 cl::desc("Force outlining cold blocks from loops."), 125 cl::init(false), cl::Hidden); 126 127 static cl::opt<bool> 128 PreciseRotationCost("precise-rotation-cost", 129 cl::desc("Model the cost of loop rotation more " 130 "precisely by using profile data."), 131 cl::init(false), cl::Hidden); 132 133 static cl::opt<bool> 134 ForcePreciseRotationCost("force-precise-rotation-cost", 135 cl::desc("Force the use of precise cost " 136 "loop rotation strategy."), 137 cl::init(false), cl::Hidden); 138 139 static cl::opt<unsigned> MisfetchCost( 140 "misfetch-cost", 141 cl::desc("Cost that models the probabilistic risk of an instruction " 142 "misfetch due to a jump comparing to falling through, whose cost " 143 "is zero."), 144 cl::init(1), cl::Hidden); 145 146 static cl::opt<unsigned> JumpInstCost("jump-inst-cost", 147 cl::desc("Cost of jump instructions."), 148 cl::init(1), cl::Hidden); 149 static cl::opt<bool> 150 TailDupPlacement("tail-dup-placement", 151 cl::desc("Perform tail duplication during placement. " 152 "Creates more fallthrough opportunites in " 153 "outline branches."), 154 cl::init(true), cl::Hidden); 155 156 static cl::opt<bool> 157 BranchFoldPlacement("branch-fold-placement", 158 cl::desc("Perform branch folding during placement. " 159 "Reduces code size."), 160 cl::init(true), cl::Hidden); 161 162 // Heuristic for tail duplication. 163 static cl::opt<unsigned> TailDupPlacementThreshold( 164 "tail-dup-placement-threshold", 165 cl::desc("Instruction cutoff for tail duplication during layout. " 166 "Tail merging during layout is forced to have a threshold " 167 "that won't conflict."), cl::init(2), 168 cl::Hidden); 169 170 // Heuristic for aggressive tail duplication. 171 static cl::opt<unsigned> TailDupPlacementAggressiveThreshold( 172 "tail-dup-placement-aggressive-threshold", 173 cl::desc("Instruction cutoff for aggressive tail duplication during " 174 "layout. Used at -O3. Tail merging during layout is forced to " 175 "have a threshold that won't conflict."), cl::init(4), 176 cl::Hidden); 177 178 // Heuristic for tail duplication. 179 static cl::opt<unsigned> TailDupPlacementPenalty( 180 "tail-dup-placement-penalty", 181 cl::desc("Cost penalty for blocks that can avoid breaking CFG by copying. " 182 "Copying can increase fallthrough, but it also increases icache " 183 "pressure. This parameter controls the penalty to account for that. " 184 "Percent as integer."), 185 cl::init(2), 186 cl::Hidden); 187 188 // Heuristic for tail duplication if profile count is used in cost model. 189 static cl::opt<unsigned> TailDupProfilePercentThreshold( 190 "tail-dup-profile-percent-threshold", 191 cl::desc("If profile count information is used in tail duplication cost " 192 "model, the gained fall through number from tail duplication " 193 "should be at least this percent of hot count."), 194 cl::init(50), cl::Hidden); 195 196 // Heuristic for triangle chains. 197 static cl::opt<unsigned> TriangleChainCount( 198 "triangle-chain-count", 199 cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the " 200 "triangle tail duplication heuristic to kick in. 0 to disable."), 201 cl::init(2), 202 cl::Hidden); 203 204 // Use case: When block layout is visualized after MBP pass, the basic blocks 205 // are labeled in layout order; meanwhile blocks could be numbered in a 206 // different order. It's hard to map between the graph and pass output. 207 // With this option on, the basic blocks are renumbered in function layout 208 // order. For debugging only. 209 static cl::opt<bool> RenumberBlocksBeforeView( 210 "renumber-blocks-before-view", 211 cl::desc( 212 "If true, basic blocks are re-numbered before MBP layout is printed " 213 "into a dot graph. Only used when a function is being printed."), 214 cl::init(false), cl::Hidden); 215 216 namespace llvm { 217 extern cl::opt<bool> EnableExtTspBlockPlacement; 218 extern cl::opt<bool> ApplyExtTspWithoutProfile; 219 extern cl::opt<unsigned> StaticLikelyProb; 220 extern cl::opt<unsigned> ProfileLikelyProb; 221 222 // Internal option used to control BFI display only after MBP pass. 223 // Defined in CodeGen/MachineBlockFrequencyInfo.cpp: 224 // -view-block-layout-with-bfi= 225 extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI; 226 227 // Command line option to specify the name of the function for CFG dump 228 // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name= 229 extern cl::opt<std::string> ViewBlockFreqFuncName; 230 } // namespace llvm 231 232 namespace { 233 234 class BlockChain; 235 236 /// Type for our function-wide basic block -> block chain mapping. 237 using BlockToChainMapType = DenseMap<const MachineBasicBlock *, BlockChain *>; 238 239 /// A chain of blocks which will be laid out contiguously. 240 /// 241 /// This is the datastructure representing a chain of consecutive blocks that 242 /// are profitable to layout together in order to maximize fallthrough 243 /// probabilities and code locality. We also can use a block chain to represent 244 /// a sequence of basic blocks which have some external (correctness) 245 /// requirement for sequential layout. 246 /// 247 /// Chains can be built around a single basic block and can be merged to grow 248 /// them. They participate in a block-to-chain mapping, which is updated 249 /// automatically as chains are merged together. 250 class BlockChain { 251 /// The sequence of blocks belonging to this chain. 252 /// 253 /// This is the sequence of blocks for a particular chain. These will be laid 254 /// out in-order within the function. 255 SmallVector<MachineBasicBlock *, 4> Blocks; 256 257 /// A handle to the function-wide basic block to block chain mapping. 258 /// 259 /// This is retained in each block chain to simplify the computation of child 260 /// block chains for SCC-formation and iteration. We store the edges to child 261 /// basic blocks, and map them back to their associated chains using this 262 /// structure. 263 BlockToChainMapType &BlockToChain; 264 265 public: 266 /// Construct a new BlockChain. 267 /// 268 /// This builds a new block chain representing a single basic block in the 269 /// function. It also registers itself as the chain that block participates 270 /// in with the BlockToChain mapping. 271 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB) 272 : Blocks(1, BB), BlockToChain(BlockToChain) { 273 assert(BB && "Cannot create a chain with a null basic block"); 274 BlockToChain[BB] = this; 275 } 276 277 /// Iterator over blocks within the chain. 278 using iterator = SmallVectorImpl<MachineBasicBlock *>::iterator; 279 using const_iterator = SmallVectorImpl<MachineBasicBlock *>::const_iterator; 280 281 /// Beginning of blocks within the chain. 282 iterator begin() { return Blocks.begin(); } 283 const_iterator begin() const { return Blocks.begin(); } 284 285 /// End of blocks within the chain. 286 iterator end() { return Blocks.end(); } 287 const_iterator end() const { return Blocks.end(); } 288 289 bool remove(MachineBasicBlock* BB) { 290 for(iterator i = begin(); i != end(); ++i) { 291 if (*i == BB) { 292 Blocks.erase(i); 293 return true; 294 } 295 } 296 return false; 297 } 298 299 /// Merge a block chain into this one. 300 /// 301 /// This routine merges a block chain into this one. It takes care of forming 302 /// a contiguous sequence of basic blocks, updating the edge list, and 303 /// updating the block -> chain mapping. It does not free or tear down the 304 /// old chain, but the old chain's block list is no longer valid. 305 void merge(MachineBasicBlock *BB, BlockChain *Chain) { 306 assert(BB && "Can't merge a null block."); 307 assert(!Blocks.empty() && "Can't merge into an empty chain."); 308 309 // Fast path in case we don't have a chain already. 310 if (!Chain) { 311 assert(!BlockToChain[BB] && 312 "Passed chain is null, but BB has entry in BlockToChain."); 313 Blocks.push_back(BB); 314 BlockToChain[BB] = this; 315 return; 316 } 317 318 assert(BB == *Chain->begin() && "Passed BB is not head of Chain."); 319 assert(Chain->begin() != Chain->end()); 320 321 // Update the incoming blocks to point to this chain, and add them to the 322 // chain structure. 323 for (MachineBasicBlock *ChainBB : *Chain) { 324 Blocks.push_back(ChainBB); 325 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain."); 326 BlockToChain[ChainBB] = this; 327 } 328 } 329 330 #ifndef NDEBUG 331 /// Dump the blocks in this chain. 332 LLVM_DUMP_METHOD void dump() { 333 for (MachineBasicBlock *MBB : *this) 334 MBB->dump(); 335 } 336 #endif // NDEBUG 337 338 /// Count of predecessors of any block within the chain which have not 339 /// yet been scheduled. In general, we will delay scheduling this chain 340 /// until those predecessors are scheduled (or we find a sufficiently good 341 /// reason to override this heuristic.) Note that when forming loop chains, 342 /// blocks outside the loop are ignored and treated as if they were already 343 /// scheduled. 344 /// 345 /// Note: This field is reinitialized multiple times - once for each loop, 346 /// and then once for the function as a whole. 347 unsigned UnscheduledPredecessors = 0; 348 }; 349 350 class MachineBlockPlacement : public MachineFunctionPass { 351 /// A type for a block filter set. 352 using BlockFilterSet = SmallSetVector<const MachineBasicBlock *, 16>; 353 354 /// Pair struct containing basic block and taildup profitability 355 struct BlockAndTailDupResult { 356 MachineBasicBlock *BB = nullptr; 357 bool ShouldTailDup; 358 }; 359 360 /// Triple struct containing edge weight and the edge. 361 struct WeightedEdge { 362 BlockFrequency Weight; 363 MachineBasicBlock *Src = nullptr; 364 MachineBasicBlock *Dest = nullptr; 365 }; 366 367 /// work lists of blocks that are ready to be laid out 368 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 369 SmallVector<MachineBasicBlock *, 16> EHPadWorkList; 370 371 /// Edges that have already been computed as optimal. 372 DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges; 373 374 /// Machine Function 375 MachineFunction *F = nullptr; 376 377 /// A handle to the branch probability pass. 378 const MachineBranchProbabilityInfo *MBPI = nullptr; 379 380 /// A handle to the function-wide block frequency pass. 381 std::unique_ptr<MBFIWrapper> MBFI; 382 383 /// A handle to the loop info. 384 MachineLoopInfo *MLI = nullptr; 385 386 /// Preferred loop exit. 387 /// Member variable for convenience. It may be removed by duplication deep 388 /// in the call stack. 389 MachineBasicBlock *PreferredLoopExit = nullptr; 390 391 /// A handle to the target's instruction info. 392 const TargetInstrInfo *TII = nullptr; 393 394 /// A handle to the target's lowering info. 395 const TargetLoweringBase *TLI = nullptr; 396 397 /// A handle to the post dominator tree. 398 MachinePostDominatorTree *MPDT = nullptr; 399 400 ProfileSummaryInfo *PSI = nullptr; 401 402 /// Duplicator used to duplicate tails during placement. 403 /// 404 /// Placement decisions can open up new tail duplication opportunities, but 405 /// since tail duplication affects placement decisions of later blocks, it 406 /// must be done inline. 407 TailDuplicator TailDup; 408 409 /// Partial tail duplication threshold. 410 BlockFrequency DupThreshold; 411 412 /// True: use block profile count to compute tail duplication cost. 413 /// False: use block frequency to compute tail duplication cost. 414 bool UseProfileCount = false; 415 416 /// Allocator and owner of BlockChain structures. 417 /// 418 /// We build BlockChains lazily while processing the loop structure of 419 /// a function. To reduce malloc traffic, we allocate them using this 420 /// slab-like allocator, and destroy them after the pass completes. An 421 /// important guarantee is that this allocator produces stable pointers to 422 /// the chains. 423 SpecificBumpPtrAllocator<BlockChain> ChainAllocator; 424 425 /// Function wide BasicBlock to BlockChain mapping. 426 /// 427 /// This mapping allows efficiently moving from any given basic block to the 428 /// BlockChain it participates in, if any. We use it to, among other things, 429 /// allow implicitly defining edges between chains as the existing edges 430 /// between basic blocks. 431 DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain; 432 433 #ifndef NDEBUG 434 /// The set of basic blocks that have terminators that cannot be fully 435 /// analyzed. These basic blocks cannot be re-ordered safely by 436 /// MachineBlockPlacement, and we must preserve physical layout of these 437 /// blocks and their successors through the pass. 438 SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits; 439 #endif 440 441 /// Get block profile count or frequency according to UseProfileCount. 442 /// The return value is used to model tail duplication cost. 443 BlockFrequency getBlockCountOrFrequency(const MachineBasicBlock *BB) { 444 if (UseProfileCount) { 445 auto Count = MBFI->getBlockProfileCount(BB); 446 if (Count) 447 return BlockFrequency(*Count); 448 else 449 return BlockFrequency(0); 450 } else 451 return MBFI->getBlockFreq(BB); 452 } 453 454 /// Scale the DupThreshold according to basic block size. 455 BlockFrequency scaleThreshold(MachineBasicBlock *BB); 456 void initDupThreshold(); 457 458 /// Decrease the UnscheduledPredecessors count for all blocks in chain, and 459 /// if the count goes to 0, add them to the appropriate work list. 460 void markChainSuccessors( 461 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB, 462 const BlockFilterSet *BlockFilter = nullptr); 463 464 /// Decrease the UnscheduledPredecessors count for a single block, and 465 /// if the count goes to 0, add them to the appropriate work list. 466 void markBlockSuccessors( 467 const BlockChain &Chain, const MachineBasicBlock *BB, 468 const MachineBasicBlock *LoopHeaderBB, 469 const BlockFilterSet *BlockFilter = nullptr); 470 471 BranchProbability 472 collectViableSuccessors( 473 const MachineBasicBlock *BB, const BlockChain &Chain, 474 const BlockFilterSet *BlockFilter, 475 SmallVector<MachineBasicBlock *, 4> &Successors); 476 bool isBestSuccessor(MachineBasicBlock *BB, MachineBasicBlock *Pred, 477 BlockFilterSet *BlockFilter); 478 void findDuplicateCandidates(SmallVectorImpl<MachineBasicBlock *> &Candidates, 479 MachineBasicBlock *BB, 480 BlockFilterSet *BlockFilter); 481 bool repeatedlyTailDuplicateBlock( 482 MachineBasicBlock *BB, MachineBasicBlock *&LPred, 483 const MachineBasicBlock *LoopHeaderBB, 484 BlockChain &Chain, BlockFilterSet *BlockFilter, 485 MachineFunction::iterator &PrevUnplacedBlockIt); 486 bool maybeTailDuplicateBlock( 487 MachineBasicBlock *BB, MachineBasicBlock *LPred, 488 BlockChain &Chain, BlockFilterSet *BlockFilter, 489 MachineFunction::iterator &PrevUnplacedBlockIt, 490 bool &DuplicatedToLPred); 491 bool hasBetterLayoutPredecessor( 492 const MachineBasicBlock *BB, const MachineBasicBlock *Succ, 493 const BlockChain &SuccChain, BranchProbability SuccProb, 494 BranchProbability RealSuccProb, const BlockChain &Chain, 495 const BlockFilterSet *BlockFilter); 496 BlockAndTailDupResult selectBestSuccessor( 497 const MachineBasicBlock *BB, const BlockChain &Chain, 498 const BlockFilterSet *BlockFilter); 499 MachineBasicBlock *selectBestCandidateBlock( 500 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList); 501 MachineBasicBlock *getFirstUnplacedBlock( 502 const BlockChain &PlacedChain, 503 MachineFunction::iterator &PrevUnplacedBlockIt, 504 const BlockFilterSet *BlockFilter); 505 506 /// Add a basic block to the work list if it is appropriate. 507 /// 508 /// If the optional parameter BlockFilter is provided, only MBB 509 /// present in the set will be added to the worklist. If nullptr 510 /// is provided, no filtering occurs. 511 void fillWorkLists(const MachineBasicBlock *MBB, 512 SmallPtrSetImpl<BlockChain *> &UpdatedPreds, 513 const BlockFilterSet *BlockFilter); 514 515 void buildChain(const MachineBasicBlock *BB, BlockChain &Chain, 516 BlockFilterSet *BlockFilter = nullptr); 517 bool canMoveBottomBlockToTop(const MachineBasicBlock *BottomBlock, 518 const MachineBasicBlock *OldTop); 519 bool hasViableTopFallthrough(const MachineBasicBlock *Top, 520 const BlockFilterSet &LoopBlockSet); 521 BlockFrequency TopFallThroughFreq(const MachineBasicBlock *Top, 522 const BlockFilterSet &LoopBlockSet); 523 BlockFrequency FallThroughGains(const MachineBasicBlock *NewTop, 524 const MachineBasicBlock *OldTop, 525 const MachineBasicBlock *ExitBB, 526 const BlockFilterSet &LoopBlockSet); 527 MachineBasicBlock *findBestLoopTopHelper(MachineBasicBlock *OldTop, 528 const MachineLoop &L, const BlockFilterSet &LoopBlockSet); 529 MachineBasicBlock *findBestLoopTop( 530 const MachineLoop &L, const BlockFilterSet &LoopBlockSet); 531 MachineBasicBlock *findBestLoopExit( 532 const MachineLoop &L, const BlockFilterSet &LoopBlockSet, 533 BlockFrequency &ExitFreq); 534 BlockFilterSet collectLoopBlockSet(const MachineLoop &L); 535 void buildLoopChains(const MachineLoop &L); 536 void rotateLoop( 537 BlockChain &LoopChain, const MachineBasicBlock *ExitingBB, 538 BlockFrequency ExitFreq, const BlockFilterSet &LoopBlockSet); 539 void rotateLoopWithProfile( 540 BlockChain &LoopChain, const MachineLoop &L, 541 const BlockFilterSet &LoopBlockSet); 542 void buildCFGChains(); 543 void optimizeBranches(); 544 void alignBlocks(); 545 /// Returns true if a block should be tail-duplicated to increase fallthrough 546 /// opportunities. 547 bool shouldTailDuplicate(MachineBasicBlock *BB); 548 /// Check the edge frequencies to see if tail duplication will increase 549 /// fallthroughs. 550 bool isProfitableToTailDup( 551 const MachineBasicBlock *BB, const MachineBasicBlock *Succ, 552 BranchProbability QProb, 553 const BlockChain &Chain, const BlockFilterSet *BlockFilter); 554 555 /// Check for a trellis layout. 556 bool isTrellis(const MachineBasicBlock *BB, 557 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs, 558 const BlockChain &Chain, const BlockFilterSet *BlockFilter); 559 560 /// Get the best successor given a trellis layout. 561 BlockAndTailDupResult getBestTrellisSuccessor( 562 const MachineBasicBlock *BB, 563 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs, 564 BranchProbability AdjustedSumProb, const BlockChain &Chain, 565 const BlockFilterSet *BlockFilter); 566 567 /// Get the best pair of non-conflicting edges. 568 static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges( 569 const MachineBasicBlock *BB, 570 MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges); 571 572 /// Returns true if a block can tail duplicate into all unplaced 573 /// predecessors. Filters based on loop. 574 bool canTailDuplicateUnplacedPreds( 575 const MachineBasicBlock *BB, MachineBasicBlock *Succ, 576 const BlockChain &Chain, const BlockFilterSet *BlockFilter); 577 578 /// Find chains of triangles to tail-duplicate where a global analysis works, 579 /// but a local analysis would not find them. 580 void precomputeTriangleChains(); 581 582 /// Apply a post-processing step optimizing block placement. 583 void applyExtTsp(); 584 585 /// Modify the existing block placement in the function and adjust all jumps. 586 void assignBlockOrder(const std::vector<const MachineBasicBlock *> &NewOrder); 587 588 /// Create a single CFG chain from the current block order. 589 void createCFGChainExtTsp(); 590 591 public: 592 static char ID; // Pass identification, replacement for typeid 593 594 MachineBlockPlacement() : MachineFunctionPass(ID) { 595 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry()); 596 } 597 598 bool runOnMachineFunction(MachineFunction &F) override; 599 600 bool allowTailDupPlacement() const { 601 assert(F); 602 return TailDupPlacement && !F->getTarget().requiresStructuredCFG(); 603 } 604 605 void getAnalysisUsage(AnalysisUsage &AU) const override { 606 AU.addRequired<MachineBranchProbabilityInfo>(); 607 AU.addRequired<MachineBlockFrequencyInfo>(); 608 if (TailDupPlacement) 609 AU.addRequired<MachinePostDominatorTree>(); 610 AU.addRequired<MachineLoopInfo>(); 611 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 612 AU.addRequired<TargetPassConfig>(); 613 MachineFunctionPass::getAnalysisUsage(AU); 614 } 615 }; 616 617 } // end anonymous namespace 618 619 char MachineBlockPlacement::ID = 0; 620 621 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID; 622 623 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE, 624 "Branch Probability Basic Block Placement", false, false) 625 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 626 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 627 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree) 628 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 629 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 630 INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE, 631 "Branch Probability Basic Block Placement", false, false) 632 633 #ifndef NDEBUG 634 /// Helper to print the name of a MBB. 635 /// 636 /// Only used by debug logging. 637 static std::string getBlockName(const MachineBasicBlock *BB) { 638 std::string Result; 639 raw_string_ostream OS(Result); 640 OS << printMBBReference(*BB); 641 OS << " ('" << BB->getName() << "')"; 642 OS.flush(); 643 return Result; 644 } 645 #endif 646 647 /// Mark a chain's successors as having one fewer preds. 648 /// 649 /// When a chain is being merged into the "placed" chain, this routine will 650 /// quickly walk the successors of each block in the chain and mark them as 651 /// having one fewer active predecessor. It also adds any successors of this 652 /// chain which reach the zero-predecessor state to the appropriate worklist. 653 void MachineBlockPlacement::markChainSuccessors( 654 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB, 655 const BlockFilterSet *BlockFilter) { 656 // Walk all the blocks in this chain, marking their successors as having 657 // a predecessor placed. 658 for (MachineBasicBlock *MBB : Chain) { 659 markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter); 660 } 661 } 662 663 /// Mark a single block's successors as having one fewer preds. 664 /// 665 /// Under normal circumstances, this is only called by markChainSuccessors, 666 /// but if a block that was to be placed is completely tail-duplicated away, 667 /// and was duplicated into the chain end, we need to redo markBlockSuccessors 668 /// for just that block. 669 void MachineBlockPlacement::markBlockSuccessors( 670 const BlockChain &Chain, const MachineBasicBlock *MBB, 671 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) { 672 // Add any successors for which this is the only un-placed in-loop 673 // predecessor to the worklist as a viable candidate for CFG-neutral 674 // placement. No subsequent placement of this block will violate the CFG 675 // shape, so we get to use heuristics to choose a favorable placement. 676 for (MachineBasicBlock *Succ : MBB->successors()) { 677 if (BlockFilter && !BlockFilter->count(Succ)) 678 continue; 679 BlockChain &SuccChain = *BlockToChain[Succ]; 680 // Disregard edges within a fixed chain, or edges to the loop header. 681 if (&Chain == &SuccChain || Succ == LoopHeaderBB) 682 continue; 683 684 // This is a cross-chain edge that is within the loop, so decrement the 685 // loop predecessor count of the destination chain. 686 if (SuccChain.UnscheduledPredecessors == 0 || 687 --SuccChain.UnscheduledPredecessors > 0) 688 continue; 689 690 auto *NewBB = *SuccChain.begin(); 691 if (NewBB->isEHPad()) 692 EHPadWorkList.push_back(NewBB); 693 else 694 BlockWorkList.push_back(NewBB); 695 } 696 } 697 698 /// This helper function collects the set of successors of block 699 /// \p BB that are allowed to be its layout successors, and return 700 /// the total branch probability of edges from \p BB to those 701 /// blocks. 702 BranchProbability MachineBlockPlacement::collectViableSuccessors( 703 const MachineBasicBlock *BB, const BlockChain &Chain, 704 const BlockFilterSet *BlockFilter, 705 SmallVector<MachineBasicBlock *, 4> &Successors) { 706 // Adjust edge probabilities by excluding edges pointing to blocks that is 707 // either not in BlockFilter or is already in the current chain. Consider the 708 // following CFG: 709 // 710 // --->A 711 // | / \ 712 // | B C 713 // | \ / \ 714 // ----D E 715 // 716 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after 717 // A->C is chosen as a fall-through, D won't be selected as a successor of C 718 // due to CFG constraint (the probability of C->D is not greater than 719 // HotProb to break topo-order). If we exclude E that is not in BlockFilter 720 // when calculating the probability of C->D, D will be selected and we 721 // will get A C D B as the layout of this loop. 722 auto AdjustedSumProb = BranchProbability::getOne(); 723 for (MachineBasicBlock *Succ : BB->successors()) { 724 bool SkipSucc = false; 725 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) { 726 SkipSucc = true; 727 } else { 728 BlockChain *SuccChain = BlockToChain[Succ]; 729 if (SuccChain == &Chain) { 730 SkipSucc = true; 731 } else if (Succ != *SuccChain->begin()) { 732 LLVM_DEBUG(dbgs() << " " << getBlockName(Succ) 733 << " -> Mid chain!\n"); 734 continue; 735 } 736 } 737 if (SkipSucc) 738 AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ); 739 else 740 Successors.push_back(Succ); 741 } 742 743 return AdjustedSumProb; 744 } 745 746 /// The helper function returns the branch probability that is adjusted 747 /// or normalized over the new total \p AdjustedSumProb. 748 static BranchProbability 749 getAdjustedProbability(BranchProbability OrigProb, 750 BranchProbability AdjustedSumProb) { 751 BranchProbability SuccProb; 752 uint32_t SuccProbN = OrigProb.getNumerator(); 753 uint32_t SuccProbD = AdjustedSumProb.getNumerator(); 754 if (SuccProbN >= SuccProbD) 755 SuccProb = BranchProbability::getOne(); 756 else 757 SuccProb = BranchProbability(SuccProbN, SuccProbD); 758 759 return SuccProb; 760 } 761 762 /// Check if \p BB has exactly the successors in \p Successors. 763 static bool 764 hasSameSuccessors(MachineBasicBlock &BB, 765 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) { 766 if (BB.succ_size() != Successors.size()) 767 return false; 768 // We don't want to count self-loops 769 if (Successors.count(&BB)) 770 return false; 771 for (MachineBasicBlock *Succ : BB.successors()) 772 if (!Successors.count(Succ)) 773 return false; 774 return true; 775 } 776 777 /// Check if a block should be tail duplicated to increase fallthrough 778 /// opportunities. 779 /// \p BB Block to check. 780 bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) { 781 // Blocks with single successors don't create additional fallthrough 782 // opportunities. Don't duplicate them. TODO: When conditional exits are 783 // analyzable, allow them to be duplicated. 784 bool IsSimple = TailDup.isSimpleBB(BB); 785 786 if (BB->succ_size() == 1) 787 return false; 788 return TailDup.shouldTailDuplicate(IsSimple, *BB); 789 } 790 791 /// Compare 2 BlockFrequency's with a small penalty for \p A. 792 /// In order to be conservative, we apply a X% penalty to account for 793 /// increased icache pressure and static heuristics. For small frequencies 794 /// we use only the numerators to improve accuracy. For simplicity, we assume the 795 /// penalty is less than 100% 796 /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere. 797 static bool greaterWithBias(BlockFrequency A, BlockFrequency B, 798 BlockFrequency EntryFreq) { 799 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100); 800 BlockFrequency Gain = A - B; 801 return (Gain / ThresholdProb) >= EntryFreq; 802 } 803 804 /// Check the edge frequencies to see if tail duplication will increase 805 /// fallthroughs. It only makes sense to call this function when 806 /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is 807 /// always locally profitable if we would have picked \p Succ without 808 /// considering duplication. 809 bool MachineBlockPlacement::isProfitableToTailDup( 810 const MachineBasicBlock *BB, const MachineBasicBlock *Succ, 811 BranchProbability QProb, 812 const BlockChain &Chain, const BlockFilterSet *BlockFilter) { 813 // We need to do a probability calculation to make sure this is profitable. 814 // First: does succ have a successor that post-dominates? This affects the 815 // calculation. The 2 relevant cases are: 816 // BB BB 817 // | \Qout | \Qout 818 // P| C |P C 819 // = C' = C' 820 // | /Qin | /Qin 821 // | / | / 822 // Succ Succ 823 // / \ | \ V 824 // U/ =V |U \ 825 // / \ = D 826 // D E | / 827 // | / 828 // |/ 829 // PDom 830 // '=' : Branch taken for that CFG edge 831 // In the second case, Placing Succ while duplicating it into C prevents the 832 // fallthrough of Succ into either D or PDom, because they now have C as an 833 // unplaced predecessor 834 835 // Start by figuring out which case we fall into 836 MachineBasicBlock *PDom = nullptr; 837 SmallVector<MachineBasicBlock *, 4> SuccSuccs; 838 // Only scan the relevant successors 839 auto AdjustedSuccSumProb = 840 collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs); 841 BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ); 842 auto BBFreq = MBFI->getBlockFreq(BB); 843 auto SuccFreq = MBFI->getBlockFreq(Succ); 844 BlockFrequency P = BBFreq * PProb; 845 BlockFrequency Qout = BBFreq * QProb; 846 BlockFrequency EntryFreq = MBFI->getEntryFreq(); 847 // If there are no more successors, it is profitable to copy, as it strictly 848 // increases fallthrough. 849 if (SuccSuccs.size() == 0) 850 return greaterWithBias(P, Qout, EntryFreq); 851 852 auto BestSuccSucc = BranchProbability::getZero(); 853 // Find the PDom or the best Succ if no PDom exists. 854 for (MachineBasicBlock *SuccSucc : SuccSuccs) { 855 auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc); 856 if (Prob > BestSuccSucc) 857 BestSuccSucc = Prob; 858 if (PDom == nullptr) 859 if (MPDT->dominates(SuccSucc, Succ)) { 860 PDom = SuccSucc; 861 break; 862 } 863 } 864 // For the comparisons, we need to know Succ's best incoming edge that isn't 865 // from BB. 866 auto SuccBestPred = BlockFrequency(0); 867 for (MachineBasicBlock *SuccPred : Succ->predecessors()) { 868 if (SuccPred == Succ || SuccPred == BB 869 || BlockToChain[SuccPred] == &Chain 870 || (BlockFilter && !BlockFilter->count(SuccPred))) 871 continue; 872 auto Freq = MBFI->getBlockFreq(SuccPred) 873 * MBPI->getEdgeProbability(SuccPred, Succ); 874 if (Freq > SuccBestPred) 875 SuccBestPred = Freq; 876 } 877 // Qin is Succ's best unplaced incoming edge that isn't BB 878 BlockFrequency Qin = SuccBestPred; 879 // If it doesn't have a post-dominating successor, here is the calculation: 880 // BB BB 881 // | \Qout | \ 882 // P| C | = 883 // = C' | C 884 // | /Qin | | 885 // | / | C' (+Succ) 886 // Succ Succ /| 887 // / \ | \/ | 888 // U/ =V | == | 889 // / \ | / \| 890 // D E D E 891 // '=' : Branch taken for that CFG edge 892 // Cost in the first case is: P + V 893 // For this calculation, we always assume P > Qout. If Qout > P 894 // The result of this function will be ignored at the caller. 895 // Let F = SuccFreq - Qin 896 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V 897 898 if (PDom == nullptr || !Succ->isSuccessor(PDom)) { 899 BranchProbability UProb = BestSuccSucc; 900 BranchProbability VProb = AdjustedSuccSumProb - UProb; 901 BlockFrequency F = SuccFreq - Qin; 902 BlockFrequency V = SuccFreq * VProb; 903 BlockFrequency QinU = std::min(Qin, F) * UProb; 904 BlockFrequency BaseCost = P + V; 905 BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb; 906 return greaterWithBias(BaseCost, DupCost, EntryFreq); 907 } 908 BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom); 909 BranchProbability VProb = AdjustedSuccSumProb - UProb; 910 BlockFrequency U = SuccFreq * UProb; 911 BlockFrequency V = SuccFreq * VProb; 912 BlockFrequency F = SuccFreq - Qin; 913 // If there is a post-dominating successor, here is the calculation: 914 // BB BB BB BB 915 // | \Qout | \ | \Qout | \ 916 // |P C | = |P C | = 917 // = C' |P C = C' |P C 918 // | /Qin | | | /Qin | | 919 // | / | C' (+Succ) | / | C' (+Succ) 920 // Succ Succ /| Succ Succ /| 921 // | \ V | \/ | | \ V | \/ | 922 // |U \ |U /\ =? |U = |U /\ | 923 // = D = = =?| | D | = =| 924 // | / |/ D | / |/ D 925 // | / | / | = | / 926 // |/ | / |/ | = 927 // Dom Dom Dom Dom 928 // '=' : Branch taken for that CFG edge 929 // The cost for taken branches in the first case is P + U 930 // Let F = SuccFreq - Qin 931 // The cost in the second case (assuming independence), given the layout: 932 // BB, Succ, (C+Succ), D, Dom or the layout: 933 // BB, Succ, D, Dom, (C+Succ) 934 // is Qout + max(F, Qin) * U + min(F, Qin) 935 // compare P + U vs Qout + P * U + Qin. 936 // 937 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ. 938 // 939 // For the 3rd case, the cost is P + 2 * V 940 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V 941 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V 942 if (UProb > AdjustedSuccSumProb / 2 && 943 !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb, 944 Chain, BlockFilter)) 945 // Cases 3 & 4 946 return greaterWithBias( 947 (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb), 948 EntryFreq); 949 // Cases 1 & 2 950 return greaterWithBias((P + U), 951 (Qout + std::min(Qin, F) * AdjustedSuccSumProb + 952 std::max(Qin, F) * UProb), 953 EntryFreq); 954 } 955 956 /// Check for a trellis layout. \p BB is the upper part of a trellis if its 957 /// successors form the lower part of a trellis. A successor set S forms the 958 /// lower part of a trellis if all of the predecessors of S are either in S or 959 /// have all of S as successors. We ignore trellises where BB doesn't have 2 960 /// successors because for fewer than 2, it's trivial, and for 3 or greater they 961 /// are very uncommon and complex to compute optimally. Allowing edges within S 962 /// is not strictly a trellis, but the same algorithm works, so we allow it. 963 bool MachineBlockPlacement::isTrellis( 964 const MachineBasicBlock *BB, 965 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs, 966 const BlockChain &Chain, const BlockFilterSet *BlockFilter) { 967 // Technically BB could form a trellis with branching factor higher than 2. 968 // But that's extremely uncommon. 969 if (BB->succ_size() != 2 || ViableSuccs.size() != 2) 970 return false; 971 972 SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(), 973 BB->succ_end()); 974 // To avoid reviewing the same predecessors twice. 975 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds; 976 977 for (MachineBasicBlock *Succ : ViableSuccs) { 978 int PredCount = 0; 979 for (auto *SuccPred : Succ->predecessors()) { 980 // Allow triangle successors, but don't count them. 981 if (Successors.count(SuccPred)) { 982 // Make sure that it is actually a triangle. 983 for (MachineBasicBlock *CheckSucc : SuccPred->successors()) 984 if (!Successors.count(CheckSucc)) 985 return false; 986 continue; 987 } 988 const BlockChain *PredChain = BlockToChain[SuccPred]; 989 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) || 990 PredChain == &Chain || PredChain == BlockToChain[Succ]) 991 continue; 992 ++PredCount; 993 // Perform the successor check only once. 994 if (!SeenPreds.insert(SuccPred).second) 995 continue; 996 if (!hasSameSuccessors(*SuccPred, Successors)) 997 return false; 998 } 999 // If one of the successors has only BB as a predecessor, it is not a 1000 // trellis. 1001 if (PredCount < 1) 1002 return false; 1003 } 1004 return true; 1005 } 1006 1007 /// Pick the highest total weight pair of edges that can both be laid out. 1008 /// The edges in \p Edges[0] are assumed to have a different destination than 1009 /// the edges in \p Edges[1]. Simple counting shows that the best pair is either 1010 /// the individual highest weight edges to the 2 different destinations, or in 1011 /// case of a conflict, one of them should be replaced with a 2nd best edge. 1012 std::pair<MachineBlockPlacement::WeightedEdge, 1013 MachineBlockPlacement::WeightedEdge> 1014 MachineBlockPlacement::getBestNonConflictingEdges( 1015 const MachineBasicBlock *BB, 1016 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>> 1017 Edges) { 1018 // Sort the edges, and then for each successor, find the best incoming 1019 // predecessor. If the best incoming predecessors aren't the same, 1020 // then that is clearly the best layout. If there is a conflict, one of the 1021 // successors will have to fallthrough from the second best predecessor. We 1022 // compare which combination is better overall. 1023 1024 // Sort for highest frequency. 1025 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; }; 1026 1027 llvm::stable_sort(Edges[0], Cmp); 1028 llvm::stable_sort(Edges[1], Cmp); 1029 auto BestA = Edges[0].begin(); 1030 auto BestB = Edges[1].begin(); 1031 // Arrange for the correct answer to be in BestA and BestB 1032 // If the 2 best edges don't conflict, the answer is already there. 1033 if (BestA->Src == BestB->Src) { 1034 // Compare the total fallthrough of (Best + Second Best) for both pairs 1035 auto SecondBestA = std::next(BestA); 1036 auto SecondBestB = std::next(BestB); 1037 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight; 1038 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight; 1039 if (BestAScore < BestBScore) 1040 BestA = SecondBestA; 1041 else 1042 BestB = SecondBestB; 1043 } 1044 // Arrange for the BB edge to be in BestA if it exists. 1045 if (BestB->Src == BB) 1046 std::swap(BestA, BestB); 1047 return std::make_pair(*BestA, *BestB); 1048 } 1049 1050 /// Get the best successor from \p BB based on \p BB being part of a trellis. 1051 /// We only handle trellises with 2 successors, so the algorithm is 1052 /// straightforward: Find the best pair of edges that don't conflict. We find 1053 /// the best incoming edge for each successor in the trellis. If those conflict, 1054 /// we consider which of them should be replaced with the second best. 1055 /// Upon return the two best edges will be in \p BestEdges. If one of the edges 1056 /// comes from \p BB, it will be in \p BestEdges[0] 1057 MachineBlockPlacement::BlockAndTailDupResult 1058 MachineBlockPlacement::getBestTrellisSuccessor( 1059 const MachineBasicBlock *BB, 1060 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs, 1061 BranchProbability AdjustedSumProb, const BlockChain &Chain, 1062 const BlockFilterSet *BlockFilter) { 1063 1064 BlockAndTailDupResult Result = {nullptr, false}; 1065 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(), 1066 BB->succ_end()); 1067 1068 // We assume size 2 because it's common. For general n, we would have to do 1069 // the Hungarian algorithm, but it's not worth the complexity because more 1070 // than 2 successors is fairly uncommon, and a trellis even more so. 1071 if (Successors.size() != 2 || ViableSuccs.size() != 2) 1072 return Result; 1073 1074 // Collect the edge frequencies of all edges that form the trellis. 1075 SmallVector<WeightedEdge, 8> Edges[2]; 1076 int SuccIndex = 0; 1077 for (auto *Succ : ViableSuccs) { 1078 for (MachineBasicBlock *SuccPred : Succ->predecessors()) { 1079 // Skip any placed predecessors that are not BB 1080 if (SuccPred != BB) 1081 if ((BlockFilter && !BlockFilter->count(SuccPred)) || 1082 BlockToChain[SuccPred] == &Chain || 1083 BlockToChain[SuccPred] == BlockToChain[Succ]) 1084 continue; 1085 BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) * 1086 MBPI->getEdgeProbability(SuccPred, Succ); 1087 Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ}); 1088 } 1089 ++SuccIndex; 1090 } 1091 1092 // Pick the best combination of 2 edges from all the edges in the trellis. 1093 WeightedEdge BestA, BestB; 1094 std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges); 1095 1096 if (BestA.Src != BB) { 1097 // If we have a trellis, and BB doesn't have the best fallthrough edges, 1098 // we shouldn't choose any successor. We've already looked and there's a 1099 // better fallthrough edge for all the successors. 1100 LLVM_DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n"); 1101 return Result; 1102 } 1103 1104 // Did we pick the triangle edge? If tail-duplication is profitable, do 1105 // that instead. Otherwise merge the triangle edge now while we know it is 1106 // optimal. 1107 if (BestA.Dest == BestB.Src) { 1108 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2 1109 // would be better. 1110 MachineBasicBlock *Succ1 = BestA.Dest; 1111 MachineBasicBlock *Succ2 = BestB.Dest; 1112 // Check to see if tail-duplication would be profitable. 1113 if (allowTailDupPlacement() && shouldTailDuplicate(Succ2) && 1114 canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) && 1115 isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1), 1116 Chain, BlockFilter)) { 1117 LLVM_DEBUG(BranchProbability Succ2Prob = getAdjustedProbability( 1118 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb); 1119 dbgs() << " Selected: " << getBlockName(Succ2) 1120 << ", probability: " << Succ2Prob 1121 << " (Tail Duplicate)\n"); 1122 Result.BB = Succ2; 1123 Result.ShouldTailDup = true; 1124 return Result; 1125 } 1126 } 1127 // We have already computed the optimal edge for the other side of the 1128 // trellis. 1129 ComputedEdges[BestB.Src] = { BestB.Dest, false }; 1130 1131 auto TrellisSucc = BestA.Dest; 1132 LLVM_DEBUG(BranchProbability SuccProb = getAdjustedProbability( 1133 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb); 1134 dbgs() << " Selected: " << getBlockName(TrellisSucc) 1135 << ", probability: " << SuccProb << " (Trellis)\n"); 1136 Result.BB = TrellisSucc; 1137 return Result; 1138 } 1139 1140 /// When the option allowTailDupPlacement() is on, this method checks if the 1141 /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated 1142 /// into all of its unplaced, unfiltered predecessors, that are not BB. 1143 bool MachineBlockPlacement::canTailDuplicateUnplacedPreds( 1144 const MachineBasicBlock *BB, MachineBasicBlock *Succ, 1145 const BlockChain &Chain, const BlockFilterSet *BlockFilter) { 1146 if (!shouldTailDuplicate(Succ)) 1147 return false; 1148 1149 // The result of canTailDuplicate. 1150 bool Duplicate = true; 1151 // Number of possible duplication. 1152 unsigned int NumDup = 0; 1153 1154 // For CFG checking. 1155 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(), 1156 BB->succ_end()); 1157 for (MachineBasicBlock *Pred : Succ->predecessors()) { 1158 // Make sure all unplaced and unfiltered predecessors can be 1159 // tail-duplicated into. 1160 // Skip any blocks that are already placed or not in this loop. 1161 if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred)) 1162 || (BlockToChain[Pred] == &Chain && !Succ->succ_empty())) 1163 continue; 1164 if (!TailDup.canTailDuplicate(Succ, Pred)) { 1165 if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors)) 1166 // This will result in a trellis after tail duplication, so we don't 1167 // need to copy Succ into this predecessor. In the presence 1168 // of a trellis tail duplication can continue to be profitable. 1169 // For example: 1170 // A A 1171 // |\ |\ 1172 // | \ | \ 1173 // | C | C+BB 1174 // | / | | 1175 // |/ | | 1176 // BB => BB | 1177 // |\ |\/| 1178 // | \ |/\| 1179 // | D | D 1180 // | / | / 1181 // |/ |/ 1182 // Succ Succ 1183 // 1184 // After BB was duplicated into C, the layout looks like the one on the 1185 // right. BB and C now have the same successors. When considering 1186 // whether Succ can be duplicated into all its unplaced predecessors, we 1187 // ignore C. 1188 // We can do this because C already has a profitable fallthrough, namely 1189 // D. TODO(iteratee): ignore sufficiently cold predecessors for 1190 // duplication and for this test. 1191 // 1192 // This allows trellises to be laid out in 2 separate chains 1193 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic 1194 // because it allows the creation of 2 fallthrough paths with links 1195 // between them, and we correctly identify the best layout for these 1196 // CFGs. We want to extend trellises that the user created in addition 1197 // to trellises created by tail-duplication, so we just look for the 1198 // CFG. 1199 continue; 1200 Duplicate = false; 1201 continue; 1202 } 1203 NumDup++; 1204 } 1205 1206 // No possible duplication in current filter set. 1207 if (NumDup == 0) 1208 return false; 1209 1210 // If profile information is available, findDuplicateCandidates can do more 1211 // precise benefit analysis. 1212 if (F->getFunction().hasProfileData()) 1213 return true; 1214 1215 // This is mainly for function exit BB. 1216 // The integrated tail duplication is really designed for increasing 1217 // fallthrough from predecessors from Succ to its successors. We may need 1218 // other machanism to handle different cases. 1219 if (Succ->succ_empty()) 1220 return true; 1221 1222 // Plus the already placed predecessor. 1223 NumDup++; 1224 1225 // If the duplication candidate has more unplaced predecessors than 1226 // successors, the extra duplication can't bring more fallthrough. 1227 // 1228 // Pred1 Pred2 Pred3 1229 // \ | / 1230 // \ | / 1231 // \ | / 1232 // Dup 1233 // / \ 1234 // / \ 1235 // Succ1 Succ2 1236 // 1237 // In this example Dup has 2 successors and 3 predecessors, duplication of Dup 1238 // can increase the fallthrough from Pred1 to Succ1 and from Pred2 to Succ2, 1239 // but the duplication into Pred3 can't increase fallthrough. 1240 // 1241 // A small number of extra duplication may not hurt too much. We need a better 1242 // heuristic to handle it. 1243 if ((NumDup > Succ->succ_size()) || !Duplicate) 1244 return false; 1245 1246 return true; 1247 } 1248 1249 /// Find chains of triangles where we believe it would be profitable to 1250 /// tail-duplicate them all, but a local analysis would not find them. 1251 /// There are 3 ways this can be profitable: 1252 /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with 1253 /// longer chains) 1254 /// 2) The chains are statically correlated. Branch probabilities have a very 1255 /// U-shaped distribution. 1256 /// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805] 1257 /// If the branches in a chain are likely to be from the same side of the 1258 /// distribution as their predecessor, but are independent at runtime, this 1259 /// transformation is profitable. (Because the cost of being wrong is a small 1260 /// fixed cost, unlike the standard triangle layout where the cost of being 1261 /// wrong scales with the # of triangles.) 1262 /// 3) The chains are dynamically correlated. If the probability that a previous 1263 /// branch was taken positively influences whether the next branch will be 1264 /// taken 1265 /// We believe that 2 and 3 are common enough to justify the small margin in 1. 1266 void MachineBlockPlacement::precomputeTriangleChains() { 1267 struct TriangleChain { 1268 std::vector<MachineBasicBlock *> Edges; 1269 1270 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst) 1271 : Edges({src, dst}) {} 1272 1273 void append(MachineBasicBlock *dst) { 1274 assert(getKey()->isSuccessor(dst) && 1275 "Attempting to append a block that is not a successor."); 1276 Edges.push_back(dst); 1277 } 1278 1279 unsigned count() const { return Edges.size() - 1; } 1280 1281 MachineBasicBlock *getKey() const { 1282 return Edges.back(); 1283 } 1284 }; 1285 1286 if (TriangleChainCount == 0) 1287 return; 1288 1289 LLVM_DEBUG(dbgs() << "Pre-computing triangle chains.\n"); 1290 // Map from last block to the chain that contains it. This allows us to extend 1291 // chains as we find new triangles. 1292 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap; 1293 for (MachineBasicBlock &BB : *F) { 1294 // If BB doesn't have 2 successors, it doesn't start a triangle. 1295 if (BB.succ_size() != 2) 1296 continue; 1297 MachineBasicBlock *PDom = nullptr; 1298 for (MachineBasicBlock *Succ : BB.successors()) { 1299 if (!MPDT->dominates(Succ, &BB)) 1300 continue; 1301 PDom = Succ; 1302 break; 1303 } 1304 // If BB doesn't have a post-dominating successor, it doesn't form a 1305 // triangle. 1306 if (PDom == nullptr) 1307 continue; 1308 // If PDom has a hint that it is low probability, skip this triangle. 1309 if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100)) 1310 continue; 1311 // If PDom isn't eligible for duplication, this isn't the kind of triangle 1312 // we're looking for. 1313 if (!shouldTailDuplicate(PDom)) 1314 continue; 1315 bool CanTailDuplicate = true; 1316 // If PDom can't tail-duplicate into it's non-BB predecessors, then this 1317 // isn't the kind of triangle we're looking for. 1318 for (MachineBasicBlock* Pred : PDom->predecessors()) { 1319 if (Pred == &BB) 1320 continue; 1321 if (!TailDup.canTailDuplicate(PDom, Pred)) { 1322 CanTailDuplicate = false; 1323 break; 1324 } 1325 } 1326 // If we can't tail-duplicate PDom to its predecessors, then skip this 1327 // triangle. 1328 if (!CanTailDuplicate) 1329 continue; 1330 1331 // Now we have an interesting triangle. Insert it if it's not part of an 1332 // existing chain. 1333 // Note: This cannot be replaced with a call insert() or emplace() because 1334 // the find key is BB, but the insert/emplace key is PDom. 1335 auto Found = TriangleChainMap.find(&BB); 1336 // If it is, remove the chain from the map, grow it, and put it back in the 1337 // map with the end as the new key. 1338 if (Found != TriangleChainMap.end()) { 1339 TriangleChain Chain = std::move(Found->second); 1340 TriangleChainMap.erase(Found); 1341 Chain.append(PDom); 1342 TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain))); 1343 } else { 1344 auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom); 1345 assert(InsertResult.second && "Block seen twice."); 1346 (void)InsertResult; 1347 } 1348 } 1349 1350 // Iterating over a DenseMap is safe here, because the only thing in the body 1351 // of the loop is inserting into another DenseMap (ComputedEdges). 1352 // ComputedEdges is never iterated, so this doesn't lead to non-determinism. 1353 for (auto &ChainPair : TriangleChainMap) { 1354 TriangleChain &Chain = ChainPair.second; 1355 // Benchmarking has shown that due to branch correlation duplicating 2 or 1356 // more triangles is profitable, despite the calculations assuming 1357 // independence. 1358 if (Chain.count() < TriangleChainCount) 1359 continue; 1360 MachineBasicBlock *dst = Chain.Edges.back(); 1361 Chain.Edges.pop_back(); 1362 for (MachineBasicBlock *src : reverse(Chain.Edges)) { 1363 LLVM_DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->" 1364 << getBlockName(dst) 1365 << " as pre-computed based on triangles.\n"); 1366 1367 auto InsertResult = ComputedEdges.insert({src, {dst, true}}); 1368 assert(InsertResult.second && "Block seen twice."); 1369 (void)InsertResult; 1370 1371 dst = src; 1372 } 1373 } 1374 } 1375 1376 // When profile is not present, return the StaticLikelyProb. 1377 // When profile is available, we need to handle the triangle-shape CFG. 1378 static BranchProbability getLayoutSuccessorProbThreshold( 1379 const MachineBasicBlock *BB) { 1380 if (!BB->getParent()->getFunction().hasProfileData()) 1381 return BranchProbability(StaticLikelyProb, 100); 1382 if (BB->succ_size() == 2) { 1383 const MachineBasicBlock *Succ1 = *BB->succ_begin(); 1384 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1); 1385 if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) { 1386 /* See case 1 below for the cost analysis. For BB->Succ to 1387 * be taken with smaller cost, the following needs to hold: 1388 * Prob(BB->Succ) > 2 * Prob(BB->Pred) 1389 * So the threshold T in the calculation below 1390 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred) 1391 * So T / (1 - T) = 2, Yielding T = 2/3 1392 * Also adding user specified branch bias, we have 1393 * T = (2/3)*(ProfileLikelyProb/50) 1394 * = (2*ProfileLikelyProb)/150) 1395 */ 1396 return BranchProbability(2 * ProfileLikelyProb, 150); 1397 } 1398 } 1399 return BranchProbability(ProfileLikelyProb, 100); 1400 } 1401 1402 /// Checks to see if the layout candidate block \p Succ has a better layout 1403 /// predecessor than \c BB. If yes, returns true. 1404 /// \p SuccProb: The probability adjusted for only remaining blocks. 1405 /// Only used for logging 1406 /// \p RealSuccProb: The un-adjusted probability. 1407 /// \p Chain: The chain that BB belongs to and Succ is being considered for. 1408 /// \p BlockFilter: if non-null, the set of blocks that make up the loop being 1409 /// considered 1410 bool MachineBlockPlacement::hasBetterLayoutPredecessor( 1411 const MachineBasicBlock *BB, const MachineBasicBlock *Succ, 1412 const BlockChain &SuccChain, BranchProbability SuccProb, 1413 BranchProbability RealSuccProb, const BlockChain &Chain, 1414 const BlockFilterSet *BlockFilter) { 1415 1416 // There isn't a better layout when there are no unscheduled predecessors. 1417 if (SuccChain.UnscheduledPredecessors == 0) 1418 return false; 1419 1420 // There are two basic scenarios here: 1421 // ------------------------------------- 1422 // Case 1: triangular shape CFG (if-then): 1423 // BB 1424 // | \ 1425 // | \ 1426 // | Pred 1427 // | / 1428 // Succ 1429 // In this case, we are evaluating whether to select edge -> Succ, e.g. 1430 // set Succ as the layout successor of BB. Picking Succ as BB's 1431 // successor breaks the CFG constraints (FIXME: define these constraints). 1432 // With this layout, Pred BB 1433 // is forced to be outlined, so the overall cost will be cost of the 1434 // branch taken from BB to Pred, plus the cost of back taken branch 1435 // from Pred to Succ, as well as the additional cost associated 1436 // with the needed unconditional jump instruction from Pred To Succ. 1437 1438 // The cost of the topological order layout is the taken branch cost 1439 // from BB to Succ, so to make BB->Succ a viable candidate, the following 1440 // must hold: 1441 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost 1442 // < freq(BB->Succ) * taken_branch_cost. 1443 // Ignoring unconditional jump cost, we get 1444 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e., 1445 // prob(BB->Succ) > 2 * prob(BB->Pred) 1446 // 1447 // When real profile data is available, we can precisely compute the 1448 // probability threshold that is needed for edge BB->Succ to be considered. 1449 // Without profile data, the heuristic requires the branch bias to be 1450 // a lot larger to make sure the signal is very strong (e.g. 80% default). 1451 // ----------------------------------------------------------------- 1452 // Case 2: diamond like CFG (if-then-else): 1453 // S 1454 // / \ 1455 // | \ 1456 // BB Pred 1457 // \ / 1458 // Succ 1459 // .. 1460 // 1461 // The current block is BB and edge BB->Succ is now being evaluated. 1462 // Note that edge S->BB was previously already selected because 1463 // prob(S->BB) > prob(S->Pred). 1464 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we 1465 // choose Pred, we will have a topological ordering as shown on the left 1466 // in the picture below. If we choose Succ, we have the solution as shown 1467 // on the right: 1468 // 1469 // topo-order: 1470 // 1471 // S----- ---S 1472 // | | | | 1473 // ---BB | | BB 1474 // | | | | 1475 // | Pred-- | Succ-- 1476 // | | | | 1477 // ---Succ ---Pred-- 1478 // 1479 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred) 1480 // = freq(S->Pred) + freq(S->BB) 1481 // 1482 // If we have profile data (i.e, branch probabilities can be trusted), the 1483 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 * 1484 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB). 1485 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which 1486 // means the cost of topological order is greater. 1487 // When profile data is not available, however, we need to be more 1488 // conservative. If the branch prediction is wrong, breaking the topo-order 1489 // will actually yield a layout with large cost. For this reason, we need 1490 // strong biased branch at block S with Prob(S->BB) in order to select 1491 // BB->Succ. This is equivalent to looking the CFG backward with backward 1492 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without 1493 // profile data). 1494 // -------------------------------------------------------------------------- 1495 // Case 3: forked diamond 1496 // S 1497 // / \ 1498 // / \ 1499 // BB Pred 1500 // | \ / | 1501 // | \ / | 1502 // | X | 1503 // | / \ | 1504 // | / \ | 1505 // S1 S2 1506 // 1507 // The current block is BB and edge BB->S1 is now being evaluated. 1508 // As above S->BB was already selected because 1509 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2). 1510 // 1511 // topo-order: 1512 // 1513 // S-------| ---S 1514 // | | | | 1515 // ---BB | | BB 1516 // | | | | 1517 // | Pred----| | S1---- 1518 // | | | | 1519 // --(S1 or S2) ---Pred-- 1520 // | 1521 // S2 1522 // 1523 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2) 1524 // + min(freq(Pred->S1), freq(Pred->S2)) 1525 // Non-topo-order cost: 1526 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2). 1527 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2)) 1528 // is 0. Then the non topo layout is better when 1529 // freq(S->Pred) < freq(BB->S1). 1530 // This is exactly what is checked below. 1531 // Note there are other shapes that apply (Pred may not be a single block, 1532 // but they all fit this general pattern.) 1533 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB); 1534 1535 // Make sure that a hot successor doesn't have a globally more 1536 // important predecessor. 1537 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb; 1538 bool BadCFGConflict = false; 1539 1540 for (MachineBasicBlock *Pred : Succ->predecessors()) { 1541 BlockChain *PredChain = BlockToChain[Pred]; 1542 if (Pred == Succ || PredChain == &SuccChain || 1543 (BlockFilter && !BlockFilter->count(Pred)) || 1544 PredChain == &Chain || Pred != *std::prev(PredChain->end()) || 1545 // This check is redundant except for look ahead. This function is 1546 // called for lookahead by isProfitableToTailDup when BB hasn't been 1547 // placed yet. 1548 (Pred == BB)) 1549 continue; 1550 // Do backward checking. 1551 // For all cases above, we need a backward checking to filter out edges that 1552 // are not 'strongly' biased. 1553 // BB Pred 1554 // \ / 1555 // Succ 1556 // We select edge BB->Succ if 1557 // freq(BB->Succ) > freq(Succ) * HotProb 1558 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) * 1559 // HotProb 1560 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb 1561 // Case 1 is covered too, because the first equation reduces to: 1562 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle) 1563 BlockFrequency PredEdgeFreq = 1564 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ); 1565 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) { 1566 BadCFGConflict = true; 1567 break; 1568 } 1569 } 1570 1571 if (BadCFGConflict) { 1572 LLVM_DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> " 1573 << SuccProb << " (prob) (non-cold CFG conflict)\n"); 1574 return true; 1575 } 1576 1577 return false; 1578 } 1579 1580 /// Select the best successor for a block. 1581 /// 1582 /// This looks across all successors of a particular block and attempts to 1583 /// select the "best" one to be the layout successor. It only considers direct 1584 /// successors which also pass the block filter. It will attempt to avoid 1585 /// breaking CFG structure, but cave and break such structures in the case of 1586 /// very hot successor edges. 1587 /// 1588 /// \returns The best successor block found, or null if none are viable, along 1589 /// with a boolean indicating if tail duplication is necessary. 1590 MachineBlockPlacement::BlockAndTailDupResult 1591 MachineBlockPlacement::selectBestSuccessor( 1592 const MachineBasicBlock *BB, const BlockChain &Chain, 1593 const BlockFilterSet *BlockFilter) { 1594 const BranchProbability HotProb(StaticLikelyProb, 100); 1595 1596 BlockAndTailDupResult BestSucc = { nullptr, false }; 1597 auto BestProb = BranchProbability::getZero(); 1598 1599 SmallVector<MachineBasicBlock *, 4> Successors; 1600 auto AdjustedSumProb = 1601 collectViableSuccessors(BB, Chain, BlockFilter, Successors); 1602 1603 LLVM_DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB) 1604 << "\n"); 1605 1606 // if we already precomputed the best successor for BB, return that if still 1607 // applicable. 1608 auto FoundEdge = ComputedEdges.find(BB); 1609 if (FoundEdge != ComputedEdges.end()) { 1610 MachineBasicBlock *Succ = FoundEdge->second.BB; 1611 ComputedEdges.erase(FoundEdge); 1612 BlockChain *SuccChain = BlockToChain[Succ]; 1613 if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) && 1614 SuccChain != &Chain && Succ == *SuccChain->begin()) 1615 return FoundEdge->second; 1616 } 1617 1618 // if BB is part of a trellis, Use the trellis to determine the optimal 1619 // fallthrough edges 1620 if (isTrellis(BB, Successors, Chain, BlockFilter)) 1621 return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain, 1622 BlockFilter); 1623 1624 // For blocks with CFG violations, we may be able to lay them out anyway with 1625 // tail-duplication. We keep this vector so we can perform the probability 1626 // calculations the minimum number of times. 1627 SmallVector<std::pair<BranchProbability, MachineBasicBlock *>, 4> 1628 DupCandidates; 1629 for (MachineBasicBlock *Succ : Successors) { 1630 auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ); 1631 BranchProbability SuccProb = 1632 getAdjustedProbability(RealSuccProb, AdjustedSumProb); 1633 1634 BlockChain &SuccChain = *BlockToChain[Succ]; 1635 // Skip the edge \c BB->Succ if block \c Succ has a better layout 1636 // predecessor that yields lower global cost. 1637 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb, 1638 Chain, BlockFilter)) { 1639 // If tail duplication would make Succ profitable, place it. 1640 if (allowTailDupPlacement() && shouldTailDuplicate(Succ)) 1641 DupCandidates.emplace_back(SuccProb, Succ); 1642 continue; 1643 } 1644 1645 LLVM_DEBUG( 1646 dbgs() << " Candidate: " << getBlockName(Succ) 1647 << ", probability: " << SuccProb 1648 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "") 1649 << "\n"); 1650 1651 if (BestSucc.BB && BestProb >= SuccProb) { 1652 LLVM_DEBUG(dbgs() << " Not the best candidate, continuing\n"); 1653 continue; 1654 } 1655 1656 LLVM_DEBUG(dbgs() << " Setting it as best candidate\n"); 1657 BestSucc.BB = Succ; 1658 BestProb = SuccProb; 1659 } 1660 // Handle the tail duplication candidates in order of decreasing probability. 1661 // Stop at the first one that is profitable. Also stop if they are less 1662 // profitable than BestSucc. Position is important because we preserve it and 1663 // prefer first best match. Here we aren't comparing in order, so we capture 1664 // the position instead. 1665 llvm::stable_sort(DupCandidates, 1666 [](std::tuple<BranchProbability, MachineBasicBlock *> L, 1667 std::tuple<BranchProbability, MachineBasicBlock *> R) { 1668 return std::get<0>(L) > std::get<0>(R); 1669 }); 1670 for (auto &Tup : DupCandidates) { 1671 BranchProbability DupProb; 1672 MachineBasicBlock *Succ; 1673 std::tie(DupProb, Succ) = Tup; 1674 if (DupProb < BestProb) 1675 break; 1676 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter) 1677 && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) { 1678 LLVM_DEBUG(dbgs() << " Candidate: " << getBlockName(Succ) 1679 << ", probability: " << DupProb 1680 << " (Tail Duplicate)\n"); 1681 BestSucc.BB = Succ; 1682 BestSucc.ShouldTailDup = true; 1683 break; 1684 } 1685 } 1686 1687 if (BestSucc.BB) 1688 LLVM_DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n"); 1689 1690 return BestSucc; 1691 } 1692 1693 /// Select the best block from a worklist. 1694 /// 1695 /// This looks through the provided worklist as a list of candidate basic 1696 /// blocks and select the most profitable one to place. The definition of 1697 /// profitable only really makes sense in the context of a loop. This returns 1698 /// the most frequently visited block in the worklist, which in the case of 1699 /// a loop, is the one most desirable to be physically close to the rest of the 1700 /// loop body in order to improve i-cache behavior. 1701 /// 1702 /// \returns The best block found, or null if none are viable. 1703 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock( 1704 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) { 1705 // Once we need to walk the worklist looking for a candidate, cleanup the 1706 // worklist of already placed entries. 1707 // FIXME: If this shows up on profiles, it could be folded (at the cost of 1708 // some code complexity) into the loop below. 1709 llvm::erase_if(WorkList, [&](MachineBasicBlock *BB) { 1710 return BlockToChain.lookup(BB) == &Chain; 1711 }); 1712 1713 if (WorkList.empty()) 1714 return nullptr; 1715 1716 bool IsEHPad = WorkList[0]->isEHPad(); 1717 1718 MachineBasicBlock *BestBlock = nullptr; 1719 BlockFrequency BestFreq; 1720 for (MachineBasicBlock *MBB : WorkList) { 1721 assert(MBB->isEHPad() == IsEHPad && 1722 "EHPad mismatch between block and work list."); 1723 1724 BlockChain &SuccChain = *BlockToChain[MBB]; 1725 if (&SuccChain == &Chain) 1726 continue; 1727 1728 assert(SuccChain.UnscheduledPredecessors == 0 && 1729 "Found CFG-violating block"); 1730 1731 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB); 1732 LLVM_DEBUG(dbgs() << " " << getBlockName(MBB) << " -> " 1733 << printBlockFreq(MBFI->getMBFI(), CandidateFreq) 1734 << " (freq)\n"); 1735 1736 // For ehpad, we layout the least probable first as to avoid jumping back 1737 // from least probable landingpads to more probable ones. 1738 // 1739 // FIXME: Using probability is probably (!) not the best way to achieve 1740 // this. We should probably have a more principled approach to layout 1741 // cleanup code. 1742 // 1743 // The goal is to get: 1744 // 1745 // +--------------------------+ 1746 // | V 1747 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume 1748 // 1749 // Rather than: 1750 // 1751 // +-------------------------------------+ 1752 // V | 1753 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup 1754 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq))) 1755 continue; 1756 1757 BestBlock = MBB; 1758 BestFreq = CandidateFreq; 1759 } 1760 1761 return BestBlock; 1762 } 1763 1764 /// Retrieve the first unplaced basic block. 1765 /// 1766 /// This routine is called when we are unable to use the CFG to walk through 1767 /// all of the basic blocks and form a chain due to unnatural loops in the CFG. 1768 /// We walk through the function's blocks in order, starting from the 1769 /// LastUnplacedBlockIt. We update this iterator on each call to avoid 1770 /// re-scanning the entire sequence on repeated calls to this routine. 1771 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock( 1772 const BlockChain &PlacedChain, 1773 MachineFunction::iterator &PrevUnplacedBlockIt, 1774 const BlockFilterSet *BlockFilter) { 1775 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E; 1776 ++I) { 1777 if (BlockFilter && !BlockFilter->count(&*I)) 1778 continue; 1779 if (BlockToChain[&*I] != &PlacedChain) { 1780 PrevUnplacedBlockIt = I; 1781 // Now select the head of the chain to which the unplaced block belongs 1782 // as the block to place. This will force the entire chain to be placed, 1783 // and satisfies the requirements of merging chains. 1784 return *BlockToChain[&*I]->begin(); 1785 } 1786 } 1787 return nullptr; 1788 } 1789 1790 void MachineBlockPlacement::fillWorkLists( 1791 const MachineBasicBlock *MBB, 1792 SmallPtrSetImpl<BlockChain *> &UpdatedPreds, 1793 const BlockFilterSet *BlockFilter = nullptr) { 1794 BlockChain &Chain = *BlockToChain[MBB]; 1795 if (!UpdatedPreds.insert(&Chain).second) 1796 return; 1797 1798 assert( 1799 Chain.UnscheduledPredecessors == 0 && 1800 "Attempting to place block with unscheduled predecessors in worklist."); 1801 for (MachineBasicBlock *ChainBB : Chain) { 1802 assert(BlockToChain[ChainBB] == &Chain && 1803 "Block in chain doesn't match BlockToChain map."); 1804 for (MachineBasicBlock *Pred : ChainBB->predecessors()) { 1805 if (BlockFilter && !BlockFilter->count(Pred)) 1806 continue; 1807 if (BlockToChain[Pred] == &Chain) 1808 continue; 1809 ++Chain.UnscheduledPredecessors; 1810 } 1811 } 1812 1813 if (Chain.UnscheduledPredecessors != 0) 1814 return; 1815 1816 MachineBasicBlock *BB = *Chain.begin(); 1817 if (BB->isEHPad()) 1818 EHPadWorkList.push_back(BB); 1819 else 1820 BlockWorkList.push_back(BB); 1821 } 1822 1823 void MachineBlockPlacement::buildChain( 1824 const MachineBasicBlock *HeadBB, BlockChain &Chain, 1825 BlockFilterSet *BlockFilter) { 1826 assert(HeadBB && "BB must not be null.\n"); 1827 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n"); 1828 MachineFunction::iterator PrevUnplacedBlockIt = F->begin(); 1829 1830 const MachineBasicBlock *LoopHeaderBB = HeadBB; 1831 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter); 1832 MachineBasicBlock *BB = *std::prev(Chain.end()); 1833 while (true) { 1834 assert(BB && "null block found at end of chain in loop."); 1835 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop."); 1836 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain."); 1837 1838 1839 // Look for the best viable successor if there is one to place immediately 1840 // after this block. 1841 auto Result = selectBestSuccessor(BB, Chain, BlockFilter); 1842 MachineBasicBlock* BestSucc = Result.BB; 1843 bool ShouldTailDup = Result.ShouldTailDup; 1844 if (allowTailDupPlacement()) 1845 ShouldTailDup |= (BestSucc && canTailDuplicateUnplacedPreds(BB, BestSucc, 1846 Chain, 1847 BlockFilter)); 1848 1849 // If an immediate successor isn't available, look for the best viable 1850 // block among those we've identified as not violating the loop's CFG at 1851 // this point. This won't be a fallthrough, but it will increase locality. 1852 if (!BestSucc) 1853 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList); 1854 if (!BestSucc) 1855 BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList); 1856 1857 if (!BestSucc) { 1858 BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter); 1859 if (!BestSucc) 1860 break; 1861 1862 LLVM_DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the " 1863 "layout successor until the CFG reduces\n"); 1864 } 1865 1866 // Placement may have changed tail duplication opportunities. 1867 // Check for that now. 1868 if (allowTailDupPlacement() && BestSucc && ShouldTailDup) { 1869 repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain, 1870 BlockFilter, PrevUnplacedBlockIt); 1871 // If the chosen successor was duplicated into BB, don't bother laying 1872 // it out, just go round the loop again with BB as the chain end. 1873 if (!BB->isSuccessor(BestSucc)) 1874 continue; 1875 } 1876 1877 // Place this block, updating the datastructures to reflect its placement. 1878 BlockChain &SuccChain = *BlockToChain[BestSucc]; 1879 // Zero out UnscheduledPredecessors for the successor we're about to merge in case 1880 // we selected a successor that didn't fit naturally into the CFG. 1881 SuccChain.UnscheduledPredecessors = 0; 1882 LLVM_DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to " 1883 << getBlockName(BestSucc) << "\n"); 1884 markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter); 1885 Chain.merge(BestSucc, &SuccChain); 1886 BB = *std::prev(Chain.end()); 1887 } 1888 1889 LLVM_DEBUG(dbgs() << "Finished forming chain for header block " 1890 << getBlockName(*Chain.begin()) << "\n"); 1891 } 1892 1893 // If bottom of block BB has only one successor OldTop, in most cases it is 1894 // profitable to move it before OldTop, except the following case: 1895 // 1896 // -->OldTop<- 1897 // | . | 1898 // | . | 1899 // | . | 1900 // ---Pred | 1901 // | | 1902 // BB----- 1903 // 1904 // If BB is moved before OldTop, Pred needs a taken branch to BB, and it can't 1905 // layout the other successor below it, so it can't reduce taken branch. 1906 // In this case we keep its original layout. 1907 bool 1908 MachineBlockPlacement::canMoveBottomBlockToTop( 1909 const MachineBasicBlock *BottomBlock, 1910 const MachineBasicBlock *OldTop) { 1911 if (BottomBlock->pred_size() != 1) 1912 return true; 1913 MachineBasicBlock *Pred = *BottomBlock->pred_begin(); 1914 if (Pred->succ_size() != 2) 1915 return true; 1916 1917 MachineBasicBlock *OtherBB = *Pred->succ_begin(); 1918 if (OtherBB == BottomBlock) 1919 OtherBB = *Pred->succ_rbegin(); 1920 if (OtherBB == OldTop) 1921 return false; 1922 1923 return true; 1924 } 1925 1926 // Find out the possible fall through frequence to the top of a loop. 1927 BlockFrequency 1928 MachineBlockPlacement::TopFallThroughFreq( 1929 const MachineBasicBlock *Top, 1930 const BlockFilterSet &LoopBlockSet) { 1931 BlockFrequency MaxFreq = BlockFrequency(0); 1932 for (MachineBasicBlock *Pred : Top->predecessors()) { 1933 BlockChain *PredChain = BlockToChain[Pred]; 1934 if (!LoopBlockSet.count(Pred) && 1935 (!PredChain || Pred == *std::prev(PredChain->end()))) { 1936 // Found a Pred block can be placed before Top. 1937 // Check if Top is the best successor of Pred. 1938 auto TopProb = MBPI->getEdgeProbability(Pred, Top); 1939 bool TopOK = true; 1940 for (MachineBasicBlock *Succ : Pred->successors()) { 1941 auto SuccProb = MBPI->getEdgeProbability(Pred, Succ); 1942 BlockChain *SuccChain = BlockToChain[Succ]; 1943 // Check if Succ can be placed after Pred. 1944 // Succ should not be in any chain, or it is the head of some chain. 1945 if (!LoopBlockSet.count(Succ) && (SuccProb > TopProb) && 1946 (!SuccChain || Succ == *SuccChain->begin())) { 1947 TopOK = false; 1948 break; 1949 } 1950 } 1951 if (TopOK) { 1952 BlockFrequency EdgeFreq = MBFI->getBlockFreq(Pred) * 1953 MBPI->getEdgeProbability(Pred, Top); 1954 if (EdgeFreq > MaxFreq) 1955 MaxFreq = EdgeFreq; 1956 } 1957 } 1958 } 1959 return MaxFreq; 1960 } 1961 1962 // Compute the fall through gains when move NewTop before OldTop. 1963 // 1964 // In following diagram, edges marked as "-" are reduced fallthrough, edges 1965 // marked as "+" are increased fallthrough, this function computes 1966 // 1967 // SUM(increased fallthrough) - SUM(decreased fallthrough) 1968 // 1969 // | 1970 // | - 1971 // V 1972 // --->OldTop 1973 // | . 1974 // | . 1975 // +| . + 1976 // | Pred ---> 1977 // | |- 1978 // | V 1979 // --- NewTop <--- 1980 // |- 1981 // V 1982 // 1983 BlockFrequency 1984 MachineBlockPlacement::FallThroughGains( 1985 const MachineBasicBlock *NewTop, 1986 const MachineBasicBlock *OldTop, 1987 const MachineBasicBlock *ExitBB, 1988 const BlockFilterSet &LoopBlockSet) { 1989 BlockFrequency FallThrough2Top = TopFallThroughFreq(OldTop, LoopBlockSet); 1990 BlockFrequency FallThrough2Exit = BlockFrequency(0); 1991 if (ExitBB) 1992 FallThrough2Exit = MBFI->getBlockFreq(NewTop) * 1993 MBPI->getEdgeProbability(NewTop, ExitBB); 1994 BlockFrequency BackEdgeFreq = MBFI->getBlockFreq(NewTop) * 1995 MBPI->getEdgeProbability(NewTop, OldTop); 1996 1997 // Find the best Pred of NewTop. 1998 MachineBasicBlock *BestPred = nullptr; 1999 BlockFrequency FallThroughFromPred = BlockFrequency(0); 2000 for (MachineBasicBlock *Pred : NewTop->predecessors()) { 2001 if (!LoopBlockSet.count(Pred)) 2002 continue; 2003 BlockChain *PredChain = BlockToChain[Pred]; 2004 if (!PredChain || Pred == *std::prev(PredChain->end())) { 2005 BlockFrequency EdgeFreq = 2006 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, NewTop); 2007 if (EdgeFreq > FallThroughFromPred) { 2008 FallThroughFromPred = EdgeFreq; 2009 BestPred = Pred; 2010 } 2011 } 2012 } 2013 2014 // If NewTop is not placed after Pred, another successor can be placed 2015 // after Pred. 2016 BlockFrequency NewFreq = BlockFrequency(0); 2017 if (BestPred) { 2018 for (MachineBasicBlock *Succ : BestPred->successors()) { 2019 if ((Succ == NewTop) || (Succ == BestPred) || !LoopBlockSet.count(Succ)) 2020 continue; 2021 if (ComputedEdges.contains(Succ)) 2022 continue; 2023 BlockChain *SuccChain = BlockToChain[Succ]; 2024 if ((SuccChain && (Succ != *SuccChain->begin())) || 2025 (SuccChain == BlockToChain[BestPred])) 2026 continue; 2027 BlockFrequency EdgeFreq = MBFI->getBlockFreq(BestPred) * 2028 MBPI->getEdgeProbability(BestPred, Succ); 2029 if (EdgeFreq > NewFreq) 2030 NewFreq = EdgeFreq; 2031 } 2032 BlockFrequency OrigEdgeFreq = MBFI->getBlockFreq(BestPred) * 2033 MBPI->getEdgeProbability(BestPred, NewTop); 2034 if (NewFreq > OrigEdgeFreq) { 2035 // If NewTop is not the best successor of Pred, then Pred doesn't 2036 // fallthrough to NewTop. So there is no FallThroughFromPred and 2037 // NewFreq. 2038 NewFreq = BlockFrequency(0); 2039 FallThroughFromPred = BlockFrequency(0); 2040 } 2041 } 2042 2043 BlockFrequency Result = BlockFrequency(0); 2044 BlockFrequency Gains = BackEdgeFreq + NewFreq; 2045 BlockFrequency Lost = 2046 FallThrough2Top + FallThrough2Exit + FallThroughFromPred; 2047 if (Gains > Lost) 2048 Result = Gains - Lost; 2049 return Result; 2050 } 2051 2052 /// Helper function of findBestLoopTop. Find the best loop top block 2053 /// from predecessors of old top. 2054 /// 2055 /// Look for a block which is strictly better than the old top for laying 2056 /// out before the old top of the loop. This looks for only two patterns: 2057 /// 2058 /// 1. a block has only one successor, the old loop top 2059 /// 2060 /// Because such a block will always result in an unconditional jump, 2061 /// rotating it in front of the old top is always profitable. 2062 /// 2063 /// 2. a block has two successors, one is old top, another is exit 2064 /// and it has more than one predecessors 2065 /// 2066 /// If it is below one of its predecessors P, only P can fall through to 2067 /// it, all other predecessors need a jump to it, and another conditional 2068 /// jump to loop header. If it is moved before loop header, all its 2069 /// predecessors jump to it, then fall through to loop header. So all its 2070 /// predecessors except P can reduce one taken branch. 2071 /// At the same time, move it before old top increases the taken branch 2072 /// to loop exit block, so the reduced taken branch will be compared with 2073 /// the increased taken branch to the loop exit block. 2074 MachineBasicBlock * 2075 MachineBlockPlacement::findBestLoopTopHelper( 2076 MachineBasicBlock *OldTop, 2077 const MachineLoop &L, 2078 const BlockFilterSet &LoopBlockSet) { 2079 // Check that the header hasn't been fused with a preheader block due to 2080 // crazy branches. If it has, we need to start with the header at the top to 2081 // prevent pulling the preheader into the loop body. 2082 BlockChain &HeaderChain = *BlockToChain[OldTop]; 2083 if (!LoopBlockSet.count(*HeaderChain.begin())) 2084 return OldTop; 2085 if (OldTop != *HeaderChain.begin()) 2086 return OldTop; 2087 2088 LLVM_DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(OldTop) 2089 << "\n"); 2090 2091 BlockFrequency BestGains = BlockFrequency(0); 2092 MachineBasicBlock *BestPred = nullptr; 2093 for (MachineBasicBlock *Pred : OldTop->predecessors()) { 2094 if (!LoopBlockSet.count(Pred)) 2095 continue; 2096 if (Pred == L.getHeader()) 2097 continue; 2098 LLVM_DEBUG(dbgs() << " old top pred: " << getBlockName(Pred) << ", has " 2099 << Pred->succ_size() << " successors, " 2100 << printBlockFreq(MBFI->getMBFI(), *Pred) << " freq\n"); 2101 if (Pred->succ_size() > 2) 2102 continue; 2103 2104 MachineBasicBlock *OtherBB = nullptr; 2105 if (Pred->succ_size() == 2) { 2106 OtherBB = *Pred->succ_begin(); 2107 if (OtherBB == OldTop) 2108 OtherBB = *Pred->succ_rbegin(); 2109 } 2110 2111 if (!canMoveBottomBlockToTop(Pred, OldTop)) 2112 continue; 2113 2114 BlockFrequency Gains = FallThroughGains(Pred, OldTop, OtherBB, 2115 LoopBlockSet); 2116 if ((Gains > BlockFrequency(0)) && 2117 (Gains > BestGains || 2118 ((Gains == BestGains) && Pred->isLayoutSuccessor(OldTop)))) { 2119 BestPred = Pred; 2120 BestGains = Gains; 2121 } 2122 } 2123 2124 // If no direct predecessor is fine, just use the loop header. 2125 if (!BestPred) { 2126 LLVM_DEBUG(dbgs() << " final top unchanged\n"); 2127 return OldTop; 2128 } 2129 2130 // Walk backwards through any straight line of predecessors. 2131 while (BestPred->pred_size() == 1 && 2132 (*BestPred->pred_begin())->succ_size() == 1 && 2133 *BestPred->pred_begin() != L.getHeader()) 2134 BestPred = *BestPred->pred_begin(); 2135 2136 LLVM_DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n"); 2137 return BestPred; 2138 } 2139 2140 /// Find the best loop top block for layout. 2141 /// 2142 /// This function iteratively calls findBestLoopTopHelper, until no new better 2143 /// BB can be found. 2144 MachineBasicBlock * 2145 MachineBlockPlacement::findBestLoopTop(const MachineLoop &L, 2146 const BlockFilterSet &LoopBlockSet) { 2147 // Placing the latch block before the header may introduce an extra branch 2148 // that skips this block the first time the loop is executed, which we want 2149 // to avoid when optimising for size. 2150 // FIXME: in theory there is a case that does not introduce a new branch, 2151 // i.e. when the layout predecessor does not fallthrough to the loop header. 2152 // In practice this never happens though: there always seems to be a preheader 2153 // that can fallthrough and that is also placed before the header. 2154 bool OptForSize = F->getFunction().hasOptSize() || 2155 llvm::shouldOptimizeForSize(L.getHeader(), PSI, MBFI.get()); 2156 if (OptForSize) 2157 return L.getHeader(); 2158 2159 MachineBasicBlock *OldTop = nullptr; 2160 MachineBasicBlock *NewTop = L.getHeader(); 2161 while (NewTop != OldTop) { 2162 OldTop = NewTop; 2163 NewTop = findBestLoopTopHelper(OldTop, L, LoopBlockSet); 2164 if (NewTop != OldTop) 2165 ComputedEdges[NewTop] = { OldTop, false }; 2166 } 2167 return NewTop; 2168 } 2169 2170 /// Find the best loop exiting block for layout. 2171 /// 2172 /// This routine implements the logic to analyze the loop looking for the best 2173 /// block to layout at the top of the loop. Typically this is done to maximize 2174 /// fallthrough opportunities. 2175 MachineBasicBlock * 2176 MachineBlockPlacement::findBestLoopExit(const MachineLoop &L, 2177 const BlockFilterSet &LoopBlockSet, 2178 BlockFrequency &ExitFreq) { 2179 // We don't want to layout the loop linearly in all cases. If the loop header 2180 // is just a normal basic block in the loop, we want to look for what block 2181 // within the loop is the best one to layout at the top. However, if the loop 2182 // header has be pre-merged into a chain due to predecessors not having 2183 // analyzable branches, *and* the predecessor it is merged with is *not* part 2184 // of the loop, rotating the header into the middle of the loop will create 2185 // a non-contiguous range of blocks which is Very Bad. So start with the 2186 // header and only rotate if safe. 2187 BlockChain &HeaderChain = *BlockToChain[L.getHeader()]; 2188 if (!LoopBlockSet.count(*HeaderChain.begin())) 2189 return nullptr; 2190 2191 BlockFrequency BestExitEdgeFreq; 2192 unsigned BestExitLoopDepth = 0; 2193 MachineBasicBlock *ExitingBB = nullptr; 2194 // If there are exits to outer loops, loop rotation can severely limit 2195 // fallthrough opportunities unless it selects such an exit. Keep a set of 2196 // blocks where rotating to exit with that block will reach an outer loop. 2197 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop; 2198 2199 LLVM_DEBUG(dbgs() << "Finding best loop exit for: " 2200 << getBlockName(L.getHeader()) << "\n"); 2201 for (MachineBasicBlock *MBB : L.getBlocks()) { 2202 BlockChain &Chain = *BlockToChain[MBB]; 2203 // Ensure that this block is at the end of a chain; otherwise it could be 2204 // mid-way through an inner loop or a successor of an unanalyzable branch. 2205 if (MBB != *std::prev(Chain.end())) 2206 continue; 2207 2208 // Now walk the successors. We need to establish whether this has a viable 2209 // exiting successor and whether it has a viable non-exiting successor. 2210 // We store the old exiting state and restore it if a viable looping 2211 // successor isn't found. 2212 MachineBasicBlock *OldExitingBB = ExitingBB; 2213 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq; 2214 bool HasLoopingSucc = false; 2215 for (MachineBasicBlock *Succ : MBB->successors()) { 2216 if (Succ->isEHPad()) 2217 continue; 2218 if (Succ == MBB) 2219 continue; 2220 BlockChain &SuccChain = *BlockToChain[Succ]; 2221 // Don't split chains, either this chain or the successor's chain. 2222 if (&Chain == &SuccChain) { 2223 LLVM_DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> " 2224 << getBlockName(Succ) << " (chain conflict)\n"); 2225 continue; 2226 } 2227 2228 auto SuccProb = MBPI->getEdgeProbability(MBB, Succ); 2229 if (LoopBlockSet.count(Succ)) { 2230 LLVM_DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> " 2231 << getBlockName(Succ) << " (" << SuccProb << ")\n"); 2232 HasLoopingSucc = true; 2233 continue; 2234 } 2235 2236 unsigned SuccLoopDepth = 0; 2237 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) { 2238 SuccLoopDepth = ExitLoop->getLoopDepth(); 2239 if (ExitLoop->contains(&L)) 2240 BlocksExitingToOuterLoop.insert(MBB); 2241 } 2242 2243 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb; 2244 LLVM_DEBUG( 2245 dbgs() << " exiting: " << getBlockName(MBB) << " -> " 2246 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (" 2247 << printBlockFreq(MBFI->getMBFI(), ExitEdgeFreq) << ")\n"); 2248 // Note that we bias this toward an existing layout successor to retain 2249 // incoming order in the absence of better information. The exit must have 2250 // a frequency higher than the current exit before we consider breaking 2251 // the layout. 2252 BranchProbability Bias(100 - ExitBlockBias, 100); 2253 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth || 2254 ExitEdgeFreq > BestExitEdgeFreq || 2255 (MBB->isLayoutSuccessor(Succ) && 2256 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) { 2257 BestExitEdgeFreq = ExitEdgeFreq; 2258 ExitingBB = MBB; 2259 } 2260 } 2261 2262 if (!HasLoopingSucc) { 2263 // Restore the old exiting state, no viable looping successor was found. 2264 ExitingBB = OldExitingBB; 2265 BestExitEdgeFreq = OldBestExitEdgeFreq; 2266 } 2267 } 2268 // Without a candidate exiting block or with only a single block in the 2269 // loop, just use the loop header to layout the loop. 2270 if (!ExitingBB) { 2271 LLVM_DEBUG( 2272 dbgs() << " No other candidate exit blocks, using loop header\n"); 2273 return nullptr; 2274 } 2275 if (L.getNumBlocks() == 1) { 2276 LLVM_DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n"); 2277 return nullptr; 2278 } 2279 2280 // Also, if we have exit blocks which lead to outer loops but didn't select 2281 // one of them as the exiting block we are rotating toward, disable loop 2282 // rotation altogether. 2283 if (!BlocksExitingToOuterLoop.empty() && 2284 !BlocksExitingToOuterLoop.count(ExitingBB)) 2285 return nullptr; 2286 2287 LLVM_DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) 2288 << "\n"); 2289 ExitFreq = BestExitEdgeFreq; 2290 return ExitingBB; 2291 } 2292 2293 /// Check if there is a fallthrough to loop header Top. 2294 /// 2295 /// 1. Look for a Pred that can be layout before Top. 2296 /// 2. Check if Top is the most possible successor of Pred. 2297 bool 2298 MachineBlockPlacement::hasViableTopFallthrough( 2299 const MachineBasicBlock *Top, 2300 const BlockFilterSet &LoopBlockSet) { 2301 for (MachineBasicBlock *Pred : Top->predecessors()) { 2302 BlockChain *PredChain = BlockToChain[Pred]; 2303 if (!LoopBlockSet.count(Pred) && 2304 (!PredChain || Pred == *std::prev(PredChain->end()))) { 2305 // Found a Pred block can be placed before Top. 2306 // Check if Top is the best successor of Pred. 2307 auto TopProb = MBPI->getEdgeProbability(Pred, Top); 2308 bool TopOK = true; 2309 for (MachineBasicBlock *Succ : Pred->successors()) { 2310 auto SuccProb = MBPI->getEdgeProbability(Pred, Succ); 2311 BlockChain *SuccChain = BlockToChain[Succ]; 2312 // Check if Succ can be placed after Pred. 2313 // Succ should not be in any chain, or it is the head of some chain. 2314 if ((!SuccChain || Succ == *SuccChain->begin()) && SuccProb > TopProb) { 2315 TopOK = false; 2316 break; 2317 } 2318 } 2319 if (TopOK) 2320 return true; 2321 } 2322 } 2323 return false; 2324 } 2325 2326 /// Attempt to rotate an exiting block to the bottom of the loop. 2327 /// 2328 /// Once we have built a chain, try to rotate it to line up the hot exit block 2329 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary 2330 /// branches. For example, if the loop has fallthrough into its header and out 2331 /// of its bottom already, don't rotate it. 2332 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain, 2333 const MachineBasicBlock *ExitingBB, 2334 BlockFrequency ExitFreq, 2335 const BlockFilterSet &LoopBlockSet) { 2336 if (!ExitingBB) 2337 return; 2338 2339 MachineBasicBlock *Top = *LoopChain.begin(); 2340 MachineBasicBlock *Bottom = *std::prev(LoopChain.end()); 2341 2342 // If ExitingBB is already the last one in a chain then nothing to do. 2343 if (Bottom == ExitingBB) 2344 return; 2345 2346 // The entry block should always be the first BB in a function. 2347 if (Top->isEntryBlock()) 2348 return; 2349 2350 bool ViableTopFallthrough = hasViableTopFallthrough(Top, LoopBlockSet); 2351 2352 // If the header has viable fallthrough, check whether the current loop 2353 // bottom is a viable exiting block. If so, bail out as rotating will 2354 // introduce an unnecessary branch. 2355 if (ViableTopFallthrough) { 2356 for (MachineBasicBlock *Succ : Bottom->successors()) { 2357 BlockChain *SuccChain = BlockToChain[Succ]; 2358 if (!LoopBlockSet.count(Succ) && 2359 (!SuccChain || Succ == *SuccChain->begin())) 2360 return; 2361 } 2362 2363 // Rotate will destroy the top fallthrough, we need to ensure the new exit 2364 // frequency is larger than top fallthrough. 2365 BlockFrequency FallThrough2Top = TopFallThroughFreq(Top, LoopBlockSet); 2366 if (FallThrough2Top >= ExitFreq) 2367 return; 2368 } 2369 2370 BlockChain::iterator ExitIt = llvm::find(LoopChain, ExitingBB); 2371 if (ExitIt == LoopChain.end()) 2372 return; 2373 2374 // Rotating a loop exit to the bottom when there is a fallthrough to top 2375 // trades the entry fallthrough for an exit fallthrough. 2376 // If there is no bottom->top edge, but the chosen exit block does have 2377 // a fallthrough, we break that fallthrough for nothing in return. 2378 2379 // Let's consider an example. We have a built chain of basic blocks 2380 // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block. 2381 // By doing a rotation we get 2382 // Bk+1, ..., Bn, B1, ..., Bk 2383 // Break of fallthrough to B1 is compensated by a fallthrough from Bk. 2384 // If we had a fallthrough Bk -> Bk+1 it is broken now. 2385 // It might be compensated by fallthrough Bn -> B1. 2386 // So we have a condition to avoid creation of extra branch by loop rotation. 2387 // All below must be true to avoid loop rotation: 2388 // If there is a fallthrough to top (B1) 2389 // There was fallthrough from chosen exit block (Bk) to next one (Bk+1) 2390 // There is no fallthrough from bottom (Bn) to top (B1). 2391 // Please note that there is no exit fallthrough from Bn because we checked it 2392 // above. 2393 if (ViableTopFallthrough) { 2394 assert(std::next(ExitIt) != LoopChain.end() && 2395 "Exit should not be last BB"); 2396 MachineBasicBlock *NextBlockInChain = *std::next(ExitIt); 2397 if (ExitingBB->isSuccessor(NextBlockInChain)) 2398 if (!Bottom->isSuccessor(Top)) 2399 return; 2400 } 2401 2402 LLVM_DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB) 2403 << " at bottom\n"); 2404 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end()); 2405 } 2406 2407 /// Attempt to rotate a loop based on profile data to reduce branch cost. 2408 /// 2409 /// With profile data, we can determine the cost in terms of missed fall through 2410 /// opportunities when rotating a loop chain and select the best rotation. 2411 /// Basically, there are three kinds of cost to consider for each rotation: 2412 /// 1. The possibly missed fall through edge (if it exists) from BB out of 2413 /// the loop to the loop header. 2414 /// 2. The possibly missed fall through edges (if they exist) from the loop 2415 /// exits to BB out of the loop. 2416 /// 3. The missed fall through edge (if it exists) from the last BB to the 2417 /// first BB in the loop chain. 2418 /// Therefore, the cost for a given rotation is the sum of costs listed above. 2419 /// We select the best rotation with the smallest cost. 2420 void MachineBlockPlacement::rotateLoopWithProfile( 2421 BlockChain &LoopChain, const MachineLoop &L, 2422 const BlockFilterSet &LoopBlockSet) { 2423 auto RotationPos = LoopChain.end(); 2424 MachineBasicBlock *ChainHeaderBB = *LoopChain.begin(); 2425 2426 // The entry block should always be the first BB in a function. 2427 if (ChainHeaderBB->isEntryBlock()) 2428 return; 2429 2430 BlockFrequency SmallestRotationCost = BlockFrequency::max(); 2431 2432 // A utility lambda that scales up a block frequency by dividing it by a 2433 // branch probability which is the reciprocal of the scale. 2434 auto ScaleBlockFrequency = [](BlockFrequency Freq, 2435 unsigned Scale) -> BlockFrequency { 2436 if (Scale == 0) 2437 return BlockFrequency(0); 2438 // Use operator / between BlockFrequency and BranchProbability to implement 2439 // saturating multiplication. 2440 return Freq / BranchProbability(1, Scale); 2441 }; 2442 2443 // Compute the cost of the missed fall-through edge to the loop header if the 2444 // chain head is not the loop header. As we only consider natural loops with 2445 // single header, this computation can be done only once. 2446 BlockFrequency HeaderFallThroughCost(0); 2447 for (auto *Pred : ChainHeaderBB->predecessors()) { 2448 BlockChain *PredChain = BlockToChain[Pred]; 2449 if (!LoopBlockSet.count(Pred) && 2450 (!PredChain || Pred == *std::prev(PredChain->end()))) { 2451 auto EdgeFreq = MBFI->getBlockFreq(Pred) * 2452 MBPI->getEdgeProbability(Pred, ChainHeaderBB); 2453 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost); 2454 // If the predecessor has only an unconditional jump to the header, we 2455 // need to consider the cost of this jump. 2456 if (Pred->succ_size() == 1) 2457 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost); 2458 HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost); 2459 } 2460 } 2461 2462 // Here we collect all exit blocks in the loop, and for each exit we find out 2463 // its hottest exit edge. For each loop rotation, we define the loop exit cost 2464 // as the sum of frequencies of exit edges we collect here, excluding the exit 2465 // edge from the tail of the loop chain. 2466 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq; 2467 for (auto *BB : LoopChain) { 2468 auto LargestExitEdgeProb = BranchProbability::getZero(); 2469 for (auto *Succ : BB->successors()) { 2470 BlockChain *SuccChain = BlockToChain[Succ]; 2471 if (!LoopBlockSet.count(Succ) && 2472 (!SuccChain || Succ == *SuccChain->begin())) { 2473 auto SuccProb = MBPI->getEdgeProbability(BB, Succ); 2474 LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb); 2475 } 2476 } 2477 if (LargestExitEdgeProb > BranchProbability::getZero()) { 2478 auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb; 2479 ExitsWithFreq.emplace_back(BB, ExitFreq); 2480 } 2481 } 2482 2483 // In this loop we iterate every block in the loop chain and calculate the 2484 // cost assuming the block is the head of the loop chain. When the loop ends, 2485 // we should have found the best candidate as the loop chain's head. 2486 for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()), 2487 EndIter = LoopChain.end(); 2488 Iter != EndIter; Iter++, TailIter++) { 2489 // TailIter is used to track the tail of the loop chain if the block we are 2490 // checking (pointed by Iter) is the head of the chain. 2491 if (TailIter == LoopChain.end()) 2492 TailIter = LoopChain.begin(); 2493 2494 auto TailBB = *TailIter; 2495 2496 // Calculate the cost by putting this BB to the top. 2497 BlockFrequency Cost = BlockFrequency(0); 2498 2499 // If the current BB is the loop header, we need to take into account the 2500 // cost of the missed fall through edge from outside of the loop to the 2501 // header. 2502 if (Iter != LoopChain.begin()) 2503 Cost += HeaderFallThroughCost; 2504 2505 // Collect the loop exit cost by summing up frequencies of all exit edges 2506 // except the one from the chain tail. 2507 for (auto &ExitWithFreq : ExitsWithFreq) 2508 if (TailBB != ExitWithFreq.first) 2509 Cost += ExitWithFreq.second; 2510 2511 // The cost of breaking the once fall-through edge from the tail to the top 2512 // of the loop chain. Here we need to consider three cases: 2513 // 1. If the tail node has only one successor, then we will get an 2514 // additional jmp instruction. So the cost here is (MisfetchCost + 2515 // JumpInstCost) * tail node frequency. 2516 // 2. If the tail node has two successors, then we may still get an 2517 // additional jmp instruction if the layout successor after the loop 2518 // chain is not its CFG successor. Note that the more frequently executed 2519 // jmp instruction will be put ahead of the other one. Assume the 2520 // frequency of those two branches are x and y, where x is the frequency 2521 // of the edge to the chain head, then the cost will be 2522 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency. 2523 // 3. If the tail node has more than two successors (this rarely happens), 2524 // we won't consider any additional cost. 2525 if (TailBB->isSuccessor(*Iter)) { 2526 auto TailBBFreq = MBFI->getBlockFreq(TailBB); 2527 if (TailBB->succ_size() == 1) 2528 Cost += ScaleBlockFrequency(TailBBFreq, MisfetchCost + JumpInstCost); 2529 else if (TailBB->succ_size() == 2) { 2530 auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter); 2531 auto TailToHeadFreq = TailBBFreq * TailToHeadProb; 2532 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2) 2533 ? TailBBFreq * TailToHeadProb.getCompl() 2534 : TailToHeadFreq; 2535 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) + 2536 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost); 2537 } 2538 } 2539 2540 LLVM_DEBUG(dbgs() << "The cost of loop rotation by making " 2541 << getBlockName(*Iter) << " to the top: " 2542 << printBlockFreq(MBFI->getMBFI(), Cost) << "\n"); 2543 2544 if (Cost < SmallestRotationCost) { 2545 SmallestRotationCost = Cost; 2546 RotationPos = Iter; 2547 } 2548 } 2549 2550 if (RotationPos != LoopChain.end()) { 2551 LLVM_DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos) 2552 << " to the top\n"); 2553 std::rotate(LoopChain.begin(), RotationPos, LoopChain.end()); 2554 } 2555 } 2556 2557 /// Collect blocks in the given loop that are to be placed. 2558 /// 2559 /// When profile data is available, exclude cold blocks from the returned set; 2560 /// otherwise, collect all blocks in the loop. 2561 MachineBlockPlacement::BlockFilterSet 2562 MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) { 2563 BlockFilterSet LoopBlockSet; 2564 2565 // Filter cold blocks off from LoopBlockSet when profile data is available. 2566 // Collect the sum of frequencies of incoming edges to the loop header from 2567 // outside. If we treat the loop as a super block, this is the frequency of 2568 // the loop. Then for each block in the loop, we calculate the ratio between 2569 // its frequency and the frequency of the loop block. When it is too small, 2570 // don't add it to the loop chain. If there are outer loops, then this block 2571 // will be merged into the first outer loop chain for which this block is not 2572 // cold anymore. This needs precise profile data and we only do this when 2573 // profile data is available. 2574 if (F->getFunction().hasProfileData() || ForceLoopColdBlock) { 2575 BlockFrequency LoopFreq(0); 2576 for (auto *LoopPred : L.getHeader()->predecessors()) 2577 if (!L.contains(LoopPred)) 2578 LoopFreq += MBFI->getBlockFreq(LoopPred) * 2579 MBPI->getEdgeProbability(LoopPred, L.getHeader()); 2580 2581 for (MachineBasicBlock *LoopBB : L.getBlocks()) { 2582 if (LoopBlockSet.count(LoopBB)) 2583 continue; 2584 auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency(); 2585 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio) 2586 continue; 2587 BlockChain *Chain = BlockToChain[LoopBB]; 2588 for (MachineBasicBlock *ChainBB : *Chain) 2589 LoopBlockSet.insert(ChainBB); 2590 } 2591 } else 2592 LoopBlockSet.insert(L.block_begin(), L.block_end()); 2593 2594 return LoopBlockSet; 2595 } 2596 2597 /// Forms basic block chains from the natural loop structures. 2598 /// 2599 /// These chains are designed to preserve the existing *structure* of the code 2600 /// as much as possible. We can then stitch the chains together in a way which 2601 /// both preserves the topological structure and minimizes taken conditional 2602 /// branches. 2603 void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) { 2604 // First recurse through any nested loops, building chains for those inner 2605 // loops. 2606 for (const MachineLoop *InnerLoop : L) 2607 buildLoopChains(*InnerLoop); 2608 2609 assert(BlockWorkList.empty() && 2610 "BlockWorkList not empty when starting to build loop chains."); 2611 assert(EHPadWorkList.empty() && 2612 "EHPadWorkList not empty when starting to build loop chains."); 2613 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L); 2614 2615 // Check if we have profile data for this function. If yes, we will rotate 2616 // this loop by modeling costs more precisely which requires the profile data 2617 // for better layout. 2618 bool RotateLoopWithProfile = 2619 ForcePreciseRotationCost || 2620 (PreciseRotationCost && F->getFunction().hasProfileData()); 2621 2622 // First check to see if there is an obviously preferable top block for the 2623 // loop. This will default to the header, but may end up as one of the 2624 // predecessors to the header if there is one which will result in strictly 2625 // fewer branches in the loop body. 2626 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet); 2627 2628 // If we selected just the header for the loop top, look for a potentially 2629 // profitable exit block in the event that rotating the loop can eliminate 2630 // branches by placing an exit edge at the bottom. 2631 // 2632 // Loops are processed innermost to uttermost, make sure we clear 2633 // PreferredLoopExit before processing a new loop. 2634 PreferredLoopExit = nullptr; 2635 BlockFrequency ExitFreq; 2636 if (!RotateLoopWithProfile && LoopTop == L.getHeader()) 2637 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet, ExitFreq); 2638 2639 BlockChain &LoopChain = *BlockToChain[LoopTop]; 2640 2641 // FIXME: This is a really lame way of walking the chains in the loop: we 2642 // walk the blocks, and use a set to prevent visiting a particular chain 2643 // twice. 2644 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 2645 assert(LoopChain.UnscheduledPredecessors == 0 && 2646 "LoopChain should not have unscheduled predecessors."); 2647 UpdatedPreds.insert(&LoopChain); 2648 2649 for (const MachineBasicBlock *LoopBB : LoopBlockSet) 2650 fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet); 2651 2652 buildChain(LoopTop, LoopChain, &LoopBlockSet); 2653 2654 if (RotateLoopWithProfile) 2655 rotateLoopWithProfile(LoopChain, L, LoopBlockSet); 2656 else 2657 rotateLoop(LoopChain, PreferredLoopExit, ExitFreq, LoopBlockSet); 2658 2659 LLVM_DEBUG({ 2660 // Crash at the end so we get all of the debugging output first. 2661 bool BadLoop = false; 2662 if (LoopChain.UnscheduledPredecessors) { 2663 BadLoop = true; 2664 dbgs() << "Loop chain contains a block without its preds placed!\n" 2665 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 2666 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"; 2667 } 2668 for (MachineBasicBlock *ChainBB : LoopChain) { 2669 dbgs() << " ... " << getBlockName(ChainBB) << "\n"; 2670 if (!LoopBlockSet.remove(ChainBB)) { 2671 // We don't mark the loop as bad here because there are real situations 2672 // where this can occur. For example, with an unanalyzable fallthrough 2673 // from a loop block to a non-loop block or vice versa. 2674 dbgs() << "Loop chain contains a block not contained by the loop!\n" 2675 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 2676 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 2677 << " Bad block: " << getBlockName(ChainBB) << "\n"; 2678 } 2679 } 2680 2681 if (!LoopBlockSet.empty()) { 2682 BadLoop = true; 2683 for (const MachineBasicBlock *LoopBB : LoopBlockSet) 2684 dbgs() << "Loop contains blocks never placed into a chain!\n" 2685 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 2686 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 2687 << " Bad block: " << getBlockName(LoopBB) << "\n"; 2688 } 2689 assert(!BadLoop && "Detected problems with the placement of this loop."); 2690 }); 2691 2692 BlockWorkList.clear(); 2693 EHPadWorkList.clear(); 2694 } 2695 2696 void MachineBlockPlacement::buildCFGChains() { 2697 // Ensure that every BB in the function has an associated chain to simplify 2698 // the assumptions of the remaining algorithm. 2699 SmallVector<MachineOperand, 4> Cond; // For analyzeBranch. 2700 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE; 2701 ++FI) { 2702 MachineBasicBlock *BB = &*FI; 2703 BlockChain *Chain = 2704 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB); 2705 // Also, merge any blocks which we cannot reason about and must preserve 2706 // the exact fallthrough behavior for. 2707 while (true) { 2708 Cond.clear(); 2709 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch. 2710 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough()) 2711 break; 2712 2713 MachineFunction::iterator NextFI = std::next(FI); 2714 MachineBasicBlock *NextBB = &*NextFI; 2715 // Ensure that the layout successor is a viable block, as we know that 2716 // fallthrough is a possibility. 2717 assert(NextFI != FE && "Can't fallthrough past the last block."); 2718 LLVM_DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: " 2719 << getBlockName(BB) << " -> " << getBlockName(NextBB) 2720 << "\n"); 2721 Chain->merge(NextBB, nullptr); 2722 #ifndef NDEBUG 2723 BlocksWithUnanalyzableExits.insert(&*BB); 2724 #endif 2725 FI = NextFI; 2726 BB = NextBB; 2727 } 2728 } 2729 2730 // Build any loop-based chains. 2731 PreferredLoopExit = nullptr; 2732 for (MachineLoop *L : *MLI) 2733 buildLoopChains(*L); 2734 2735 assert(BlockWorkList.empty() && 2736 "BlockWorkList should be empty before building final chain."); 2737 assert(EHPadWorkList.empty() && 2738 "EHPadWorkList should be empty before building final chain."); 2739 2740 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 2741 for (MachineBasicBlock &MBB : *F) 2742 fillWorkLists(&MBB, UpdatedPreds); 2743 2744 BlockChain &FunctionChain = *BlockToChain[&F->front()]; 2745 buildChain(&F->front(), FunctionChain); 2746 2747 #ifndef NDEBUG 2748 using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>; 2749 #endif 2750 LLVM_DEBUG({ 2751 // Crash at the end so we get all of the debugging output first. 2752 bool BadFunc = false; 2753 FunctionBlockSetType FunctionBlockSet; 2754 for (MachineBasicBlock &MBB : *F) 2755 FunctionBlockSet.insert(&MBB); 2756 2757 for (MachineBasicBlock *ChainBB : FunctionChain) 2758 if (!FunctionBlockSet.erase(ChainBB)) { 2759 BadFunc = true; 2760 dbgs() << "Function chain contains a block not in the function!\n" 2761 << " Bad block: " << getBlockName(ChainBB) << "\n"; 2762 } 2763 2764 if (!FunctionBlockSet.empty()) { 2765 BadFunc = true; 2766 for (MachineBasicBlock *RemainingBB : FunctionBlockSet) 2767 dbgs() << "Function contains blocks never placed into a chain!\n" 2768 << " Bad block: " << getBlockName(RemainingBB) << "\n"; 2769 } 2770 assert(!BadFunc && "Detected problems with the block placement."); 2771 }); 2772 2773 // Remember original layout ordering, so we can update terminators after 2774 // reordering to point to the original layout successor. 2775 SmallVector<MachineBasicBlock *, 4> OriginalLayoutSuccessors( 2776 F->getNumBlockIDs()); 2777 { 2778 MachineBasicBlock *LastMBB = nullptr; 2779 for (auto &MBB : *F) { 2780 if (LastMBB != nullptr) 2781 OriginalLayoutSuccessors[LastMBB->getNumber()] = &MBB; 2782 LastMBB = &MBB; 2783 } 2784 OriginalLayoutSuccessors[F->back().getNumber()] = nullptr; 2785 } 2786 2787 // Splice the blocks into place. 2788 MachineFunction::iterator InsertPos = F->begin(); 2789 LLVM_DEBUG(dbgs() << "[MBP] Function: " << F->getName() << "\n"); 2790 for (MachineBasicBlock *ChainBB : FunctionChain) { 2791 LLVM_DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain " 2792 : " ... ") 2793 << getBlockName(ChainBB) << "\n"); 2794 if (InsertPos != MachineFunction::iterator(ChainBB)) 2795 F->splice(InsertPos, ChainBB); 2796 else 2797 ++InsertPos; 2798 2799 // Update the terminator of the previous block. 2800 if (ChainBB == *FunctionChain.begin()) 2801 continue; 2802 MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB)); 2803 2804 // FIXME: It would be awesome of updateTerminator would just return rather 2805 // than assert when the branch cannot be analyzed in order to remove this 2806 // boiler plate. 2807 Cond.clear(); 2808 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch. 2809 2810 #ifndef NDEBUG 2811 if (!BlocksWithUnanalyzableExits.count(PrevBB)) { 2812 // Given the exact block placement we chose, we may actually not _need_ to 2813 // be able to edit PrevBB's terminator sequence, but not being _able_ to 2814 // do that at this point is a bug. 2815 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) || 2816 !PrevBB->canFallThrough()) && 2817 "Unexpected block with un-analyzable fallthrough!"); 2818 Cond.clear(); 2819 TBB = FBB = nullptr; 2820 } 2821 #endif 2822 2823 // The "PrevBB" is not yet updated to reflect current code layout, so, 2824 // o. it may fall-through to a block without explicit "goto" instruction 2825 // before layout, and no longer fall-through it after layout; or 2826 // o. just opposite. 2827 // 2828 // analyzeBranch() may return erroneous value for FBB when these two 2829 // situations take place. For the first scenario FBB is mistakenly set NULL; 2830 // for the 2nd scenario, the FBB, which is expected to be NULL, is 2831 // mistakenly pointing to "*BI". 2832 // Thus, if the future change needs to use FBB before the layout is set, it 2833 // has to correct FBB first by using the code similar to the following: 2834 // 2835 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) { 2836 // PrevBB->updateTerminator(); 2837 // Cond.clear(); 2838 // TBB = FBB = nullptr; 2839 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) { 2840 // // FIXME: This should never take place. 2841 // TBB = FBB = nullptr; 2842 // } 2843 // } 2844 if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) { 2845 PrevBB->updateTerminator(OriginalLayoutSuccessors[PrevBB->getNumber()]); 2846 } 2847 } 2848 2849 // Fixup the last block. 2850 Cond.clear(); 2851 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch. 2852 if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond)) { 2853 MachineBasicBlock *PrevBB = &F->back(); 2854 PrevBB->updateTerminator(OriginalLayoutSuccessors[PrevBB->getNumber()]); 2855 } 2856 2857 BlockWorkList.clear(); 2858 EHPadWorkList.clear(); 2859 } 2860 2861 void MachineBlockPlacement::optimizeBranches() { 2862 BlockChain &FunctionChain = *BlockToChain[&F->front()]; 2863 SmallVector<MachineOperand, 4> Cond; // For analyzeBranch. 2864 2865 // Now that all the basic blocks in the chain have the proper layout, 2866 // make a final call to analyzeBranch with AllowModify set. 2867 // Indeed, the target may be able to optimize the branches in a way we 2868 // cannot because all branches may not be analyzable. 2869 // E.g., the target may be able to remove an unconditional branch to 2870 // a fallthrough when it occurs after predicated terminators. 2871 for (MachineBasicBlock *ChainBB : FunctionChain) { 2872 Cond.clear(); 2873 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch. 2874 if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) { 2875 // If PrevBB has a two-way branch, try to re-order the branches 2876 // such that we branch to the successor with higher probability first. 2877 if (TBB && !Cond.empty() && FBB && 2878 MBPI->getEdgeProbability(ChainBB, FBB) > 2879 MBPI->getEdgeProbability(ChainBB, TBB) && 2880 !TII->reverseBranchCondition(Cond)) { 2881 LLVM_DEBUG(dbgs() << "Reverse order of the two branches: " 2882 << getBlockName(ChainBB) << "\n"); 2883 LLVM_DEBUG(dbgs() << " Edge probability: " 2884 << MBPI->getEdgeProbability(ChainBB, FBB) << " vs " 2885 << MBPI->getEdgeProbability(ChainBB, TBB) << "\n"); 2886 DebugLoc dl; // FIXME: this is nowhere 2887 TII->removeBranch(*ChainBB); 2888 TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl); 2889 } 2890 } 2891 } 2892 } 2893 2894 void MachineBlockPlacement::alignBlocks() { 2895 // Walk through the backedges of the function now that we have fully laid out 2896 // the basic blocks and align the destination of each backedge. We don't rely 2897 // exclusively on the loop info here so that we can align backedges in 2898 // unnatural CFGs and backedges that were introduced purely because of the 2899 // loop rotations done during this layout pass. 2900 if (F->getFunction().hasMinSize() || 2901 (F->getFunction().hasOptSize() && !TLI->alignLoopsWithOptSize())) 2902 return; 2903 BlockChain &FunctionChain = *BlockToChain[&F->front()]; 2904 if (FunctionChain.begin() == FunctionChain.end()) 2905 return; // Empty chain. 2906 2907 const BranchProbability ColdProb(1, 5); // 20% 2908 BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front()); 2909 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb; 2910 for (MachineBasicBlock *ChainBB : FunctionChain) { 2911 if (ChainBB == *FunctionChain.begin()) 2912 continue; 2913 2914 // Don't align non-looping basic blocks. These are unlikely to execute 2915 // enough times to matter in practice. Note that we'll still handle 2916 // unnatural CFGs inside of a natural outer loop (the common case) and 2917 // rotated loops. 2918 MachineLoop *L = MLI->getLoopFor(ChainBB); 2919 if (!L) 2920 continue; 2921 2922 const Align TLIAlign = TLI->getPrefLoopAlignment(L); 2923 unsigned MDAlign = 1; 2924 MDNode *LoopID = L->getLoopID(); 2925 if (LoopID) { 2926 for (unsigned I = 1, E = LoopID->getNumOperands(); I < E; ++I) { 2927 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(I)); 2928 if (MD == nullptr) 2929 continue; 2930 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 2931 if (S == nullptr) 2932 continue; 2933 if (S->getString() == "llvm.loop.align") { 2934 assert(MD->getNumOperands() == 2 && 2935 "per-loop align metadata should have two operands."); 2936 MDAlign = 2937 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 2938 assert(MDAlign >= 1 && "per-loop align value must be positive."); 2939 } 2940 } 2941 } 2942 2943 // Use max of the TLIAlign and MDAlign 2944 const Align LoopAlign = std::max(TLIAlign, Align(MDAlign)); 2945 if (LoopAlign == 1) 2946 continue; // Don't care about loop alignment. 2947 2948 // If the block is cold relative to the function entry don't waste space 2949 // aligning it. 2950 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB); 2951 if (Freq < WeightedEntryFreq) 2952 continue; 2953 2954 // If the block is cold relative to its loop header, don't align it 2955 // regardless of what edges into the block exist. 2956 MachineBasicBlock *LoopHeader = L->getHeader(); 2957 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader); 2958 if (Freq < (LoopHeaderFreq * ColdProb)) 2959 continue; 2960 2961 // If the global profiles indicates so, don't align it. 2962 if (llvm::shouldOptimizeForSize(ChainBB, PSI, MBFI.get()) && 2963 !TLI->alignLoopsWithOptSize()) 2964 continue; 2965 2966 // Check for the existence of a non-layout predecessor which would benefit 2967 // from aligning this block. 2968 MachineBasicBlock *LayoutPred = 2969 &*std::prev(MachineFunction::iterator(ChainBB)); 2970 2971 auto DetermineMaxAlignmentPadding = [&]() { 2972 // Set the maximum bytes allowed to be emitted for alignment. 2973 unsigned MaxBytes; 2974 if (MaxBytesForAlignmentOverride.getNumOccurrences() > 0) 2975 MaxBytes = MaxBytesForAlignmentOverride; 2976 else 2977 MaxBytes = TLI->getMaxPermittedBytesForAlignment(ChainBB); 2978 ChainBB->setMaxBytesForAlignment(MaxBytes); 2979 }; 2980 2981 // Force alignment if all the predecessors are jumps. We already checked 2982 // that the block isn't cold above. 2983 if (!LayoutPred->isSuccessor(ChainBB)) { 2984 ChainBB->setAlignment(LoopAlign); 2985 DetermineMaxAlignmentPadding(); 2986 continue; 2987 } 2988 2989 // Align this block if the layout predecessor's edge into this block is 2990 // cold relative to the block. When this is true, other predecessors make up 2991 // all of the hot entries into the block and thus alignment is likely to be 2992 // important. 2993 BranchProbability LayoutProb = 2994 MBPI->getEdgeProbability(LayoutPred, ChainBB); 2995 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb; 2996 if (LayoutEdgeFreq <= (Freq * ColdProb)) { 2997 ChainBB->setAlignment(LoopAlign); 2998 DetermineMaxAlignmentPadding(); 2999 } 3000 } 3001 } 3002 3003 /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if 3004 /// it was duplicated into its chain predecessor and removed. 3005 /// \p BB - Basic block that may be duplicated. 3006 /// 3007 /// \p LPred - Chosen layout predecessor of \p BB. 3008 /// Updated to be the chain end if LPred is removed. 3009 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong. 3010 /// \p BlockFilter - Set of blocks that belong to the loop being laid out. 3011 /// Used to identify which blocks to update predecessor 3012 /// counts. 3013 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was 3014 /// chosen in the given order due to unnatural CFG 3015 /// only needed if \p BB is removed and 3016 /// \p PrevUnplacedBlockIt pointed to \p BB. 3017 /// @return true if \p BB was removed. 3018 bool MachineBlockPlacement::repeatedlyTailDuplicateBlock( 3019 MachineBasicBlock *BB, MachineBasicBlock *&LPred, 3020 const MachineBasicBlock *LoopHeaderBB, 3021 BlockChain &Chain, BlockFilterSet *BlockFilter, 3022 MachineFunction::iterator &PrevUnplacedBlockIt) { 3023 bool Removed, DuplicatedToLPred; 3024 bool DuplicatedToOriginalLPred; 3025 Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter, 3026 PrevUnplacedBlockIt, 3027 DuplicatedToLPred); 3028 if (!Removed) 3029 return false; 3030 DuplicatedToOriginalLPred = DuplicatedToLPred; 3031 // Iteratively try to duplicate again. It can happen that a block that is 3032 // duplicated into is still small enough to be duplicated again. 3033 // No need to call markBlockSuccessors in this case, as the blocks being 3034 // duplicated from here on are already scheduled. 3035 while (DuplicatedToLPred && Removed) { 3036 MachineBasicBlock *DupBB, *DupPred; 3037 // The removal callback causes Chain.end() to be updated when a block is 3038 // removed. On the first pass through the loop, the chain end should be the 3039 // same as it was on function entry. On subsequent passes, because we are 3040 // duplicating the block at the end of the chain, if it is removed the 3041 // chain will have shrunk by one block. 3042 BlockChain::iterator ChainEnd = Chain.end(); 3043 DupBB = *(--ChainEnd); 3044 // Now try to duplicate again. 3045 if (ChainEnd == Chain.begin()) 3046 break; 3047 DupPred = *std::prev(ChainEnd); 3048 Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter, 3049 PrevUnplacedBlockIt, 3050 DuplicatedToLPred); 3051 } 3052 // If BB was duplicated into LPred, it is now scheduled. But because it was 3053 // removed, markChainSuccessors won't be called for its chain. Instead we 3054 // call markBlockSuccessors for LPred to achieve the same effect. This must go 3055 // at the end because repeating the tail duplication can increase the number 3056 // of unscheduled predecessors. 3057 LPred = *std::prev(Chain.end()); 3058 if (DuplicatedToOriginalLPred) 3059 markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter); 3060 return true; 3061 } 3062 3063 /// Tail duplicate \p BB into (some) predecessors if profitable. 3064 /// \p BB - Basic block that may be duplicated 3065 /// \p LPred - Chosen layout predecessor of \p BB 3066 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong. 3067 /// \p BlockFilter - Set of blocks that belong to the loop being laid out. 3068 /// Used to identify which blocks to update predecessor 3069 /// counts. 3070 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was 3071 /// chosen in the given order due to unnatural CFG 3072 /// only needed if \p BB is removed and 3073 /// \p PrevUnplacedBlockIt pointed to \p BB. 3074 /// \p DuplicatedToLPred - True if the block was duplicated into LPred. 3075 /// \return - True if the block was duplicated into all preds and removed. 3076 bool MachineBlockPlacement::maybeTailDuplicateBlock( 3077 MachineBasicBlock *BB, MachineBasicBlock *LPred, 3078 BlockChain &Chain, BlockFilterSet *BlockFilter, 3079 MachineFunction::iterator &PrevUnplacedBlockIt, 3080 bool &DuplicatedToLPred) { 3081 DuplicatedToLPred = false; 3082 if (!shouldTailDuplicate(BB)) 3083 return false; 3084 3085 LLVM_DEBUG(dbgs() << "Redoing tail duplication for Succ#" << BB->getNumber() 3086 << "\n"); 3087 3088 // This has to be a callback because none of it can be done after 3089 // BB is deleted. 3090 bool Removed = false; 3091 auto RemovalCallback = 3092 [&](MachineBasicBlock *RemBB) { 3093 // Signal to outer function 3094 Removed = true; 3095 3096 // Conservative default. 3097 bool InWorkList = true; 3098 // Remove from the Chain and Chain Map 3099 if (BlockToChain.count(RemBB)) { 3100 BlockChain *Chain = BlockToChain[RemBB]; 3101 InWorkList = Chain->UnscheduledPredecessors == 0; 3102 Chain->remove(RemBB); 3103 BlockToChain.erase(RemBB); 3104 } 3105 3106 // Handle the unplaced block iterator 3107 if (&(*PrevUnplacedBlockIt) == RemBB) { 3108 PrevUnplacedBlockIt++; 3109 } 3110 3111 // Handle the Work Lists 3112 if (InWorkList) { 3113 SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList; 3114 if (RemBB->isEHPad()) 3115 RemoveList = EHPadWorkList; 3116 llvm::erase(RemoveList, RemBB); 3117 } 3118 3119 // Handle the filter set 3120 if (BlockFilter) { 3121 BlockFilter->remove(RemBB); 3122 } 3123 3124 // Remove the block from loop info. 3125 MLI->removeBlock(RemBB); 3126 if (RemBB == PreferredLoopExit) 3127 PreferredLoopExit = nullptr; 3128 3129 LLVM_DEBUG(dbgs() << "TailDuplicator deleted block: " 3130 << getBlockName(RemBB) << "\n"); 3131 }; 3132 auto RemovalCallbackRef = 3133 function_ref<void(MachineBasicBlock*)>(RemovalCallback); 3134 3135 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds; 3136 bool IsSimple = TailDup.isSimpleBB(BB); 3137 SmallVector<MachineBasicBlock *, 8> CandidatePreds; 3138 SmallVectorImpl<MachineBasicBlock *> *CandidatePtr = nullptr; 3139 if (F->getFunction().hasProfileData()) { 3140 // We can do partial duplication with precise profile information. 3141 findDuplicateCandidates(CandidatePreds, BB, BlockFilter); 3142 if (CandidatePreds.size() == 0) 3143 return false; 3144 if (CandidatePreds.size() < BB->pred_size()) 3145 CandidatePtr = &CandidatePreds; 3146 } 3147 TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred, &DuplicatedPreds, 3148 &RemovalCallbackRef, CandidatePtr); 3149 3150 // Update UnscheduledPredecessors to reflect tail-duplication. 3151 DuplicatedToLPred = false; 3152 for (MachineBasicBlock *Pred : DuplicatedPreds) { 3153 // We're only looking for unscheduled predecessors that match the filter. 3154 BlockChain* PredChain = BlockToChain[Pred]; 3155 if (Pred == LPred) 3156 DuplicatedToLPred = true; 3157 if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred)) 3158 || PredChain == &Chain) 3159 continue; 3160 for (MachineBasicBlock *NewSucc : Pred->successors()) { 3161 if (BlockFilter && !BlockFilter->count(NewSucc)) 3162 continue; 3163 BlockChain *NewChain = BlockToChain[NewSucc]; 3164 if (NewChain != &Chain && NewChain != PredChain) 3165 NewChain->UnscheduledPredecessors++; 3166 } 3167 } 3168 return Removed; 3169 } 3170 3171 // Count the number of actual machine instructions. 3172 static uint64_t countMBBInstruction(MachineBasicBlock *MBB) { 3173 uint64_t InstrCount = 0; 3174 for (MachineInstr &MI : *MBB) { 3175 if (!MI.isPHI() && !MI.isMetaInstruction()) 3176 InstrCount += 1; 3177 } 3178 return InstrCount; 3179 } 3180 3181 // The size cost of duplication is the instruction size of the duplicated block. 3182 // So we should scale the threshold accordingly. But the instruction size is not 3183 // available on all targets, so we use the number of instructions instead. 3184 BlockFrequency MachineBlockPlacement::scaleThreshold(MachineBasicBlock *BB) { 3185 return BlockFrequency(DupThreshold.getFrequency() * countMBBInstruction(BB)); 3186 } 3187 3188 // Returns true if BB is Pred's best successor. 3189 bool MachineBlockPlacement::isBestSuccessor(MachineBasicBlock *BB, 3190 MachineBasicBlock *Pred, 3191 BlockFilterSet *BlockFilter) { 3192 if (BB == Pred) 3193 return false; 3194 if (BlockFilter && !BlockFilter->count(Pred)) 3195 return false; 3196 BlockChain *PredChain = BlockToChain[Pred]; 3197 if (PredChain && (Pred != *std::prev(PredChain->end()))) 3198 return false; 3199 3200 // Find the successor with largest probability excluding BB. 3201 BranchProbability BestProb = BranchProbability::getZero(); 3202 for (MachineBasicBlock *Succ : Pred->successors()) 3203 if (Succ != BB) { 3204 if (BlockFilter && !BlockFilter->count(Succ)) 3205 continue; 3206 BlockChain *SuccChain = BlockToChain[Succ]; 3207 if (SuccChain && (Succ != *SuccChain->begin())) 3208 continue; 3209 BranchProbability SuccProb = MBPI->getEdgeProbability(Pred, Succ); 3210 if (SuccProb > BestProb) 3211 BestProb = SuccProb; 3212 } 3213 3214 BranchProbability BBProb = MBPI->getEdgeProbability(Pred, BB); 3215 if (BBProb <= BestProb) 3216 return false; 3217 3218 // Compute the number of reduced taken branches if Pred falls through to BB 3219 // instead of another successor. Then compare it with threshold. 3220 BlockFrequency PredFreq = getBlockCountOrFrequency(Pred); 3221 BlockFrequency Gain = PredFreq * (BBProb - BestProb); 3222 return Gain > scaleThreshold(BB); 3223 } 3224 3225 // Find out the predecessors of BB and BB can be beneficially duplicated into 3226 // them. 3227 void MachineBlockPlacement::findDuplicateCandidates( 3228 SmallVectorImpl<MachineBasicBlock *> &Candidates, 3229 MachineBasicBlock *BB, 3230 BlockFilterSet *BlockFilter) { 3231 MachineBasicBlock *Fallthrough = nullptr; 3232 BranchProbability DefaultBranchProb = BranchProbability::getZero(); 3233 BlockFrequency BBDupThreshold(scaleThreshold(BB)); 3234 SmallVector<MachineBasicBlock *, 8> Preds(BB->predecessors()); 3235 SmallVector<MachineBasicBlock *, 8> Succs(BB->successors()); 3236 3237 // Sort for highest frequency. 3238 auto CmpSucc = [&](MachineBasicBlock *A, MachineBasicBlock *B) { 3239 return MBPI->getEdgeProbability(BB, A) > MBPI->getEdgeProbability(BB, B); 3240 }; 3241 auto CmpPred = [&](MachineBasicBlock *A, MachineBasicBlock *B) { 3242 return MBFI->getBlockFreq(A) > MBFI->getBlockFreq(B); 3243 }; 3244 llvm::stable_sort(Succs, CmpSucc); 3245 llvm::stable_sort(Preds, CmpPred); 3246 3247 auto SuccIt = Succs.begin(); 3248 if (SuccIt != Succs.end()) { 3249 DefaultBranchProb = MBPI->getEdgeProbability(BB, *SuccIt).getCompl(); 3250 } 3251 3252 // For each predecessors of BB, compute the benefit of duplicating BB, 3253 // if it is larger than the threshold, add it into Candidates. 3254 // 3255 // If we have following control flow. 3256 // 3257 // PB1 PB2 PB3 PB4 3258 // \ | / /\ 3259 // \ | / / \ 3260 // \ |/ / \ 3261 // BB----/ OB 3262 // /\ 3263 // / \ 3264 // SB1 SB2 3265 // 3266 // And it can be partially duplicated as 3267 // 3268 // PB2+BB 3269 // | PB1 PB3 PB4 3270 // | | / /\ 3271 // | | / / \ 3272 // | |/ / \ 3273 // | BB----/ OB 3274 // |\ /| 3275 // | X | 3276 // |/ \| 3277 // SB2 SB1 3278 // 3279 // The benefit of duplicating into a predecessor is defined as 3280 // Orig_taken_branch - Duplicated_taken_branch 3281 // 3282 // The Orig_taken_branch is computed with the assumption that predecessor 3283 // jumps to BB and the most possible successor is laid out after BB. 3284 // 3285 // The Duplicated_taken_branch is computed with the assumption that BB is 3286 // duplicated into PB, and one successor is layout after it (SB1 for PB1 and 3287 // SB2 for PB2 in our case). If there is no available successor, the combined 3288 // block jumps to all BB's successor, like PB3 in this example. 3289 // 3290 // If a predecessor has multiple successors, so BB can't be duplicated into 3291 // it. But it can beneficially fall through to BB, and duplicate BB into other 3292 // predecessors. 3293 for (MachineBasicBlock *Pred : Preds) { 3294 BlockFrequency PredFreq = getBlockCountOrFrequency(Pred); 3295 3296 if (!TailDup.canTailDuplicate(BB, Pred)) { 3297 // BB can't be duplicated into Pred, but it is possible to be layout 3298 // below Pred. 3299 if (!Fallthrough && isBestSuccessor(BB, Pred, BlockFilter)) { 3300 Fallthrough = Pred; 3301 if (SuccIt != Succs.end()) 3302 SuccIt++; 3303 } 3304 continue; 3305 } 3306 3307 BlockFrequency OrigCost = PredFreq + PredFreq * DefaultBranchProb; 3308 BlockFrequency DupCost; 3309 if (SuccIt == Succs.end()) { 3310 // Jump to all successors; 3311 if (Succs.size() > 0) 3312 DupCost += PredFreq; 3313 } else { 3314 // Fallthrough to *SuccIt, jump to all other successors; 3315 DupCost += PredFreq; 3316 DupCost -= PredFreq * MBPI->getEdgeProbability(BB, *SuccIt); 3317 } 3318 3319 assert(OrigCost >= DupCost); 3320 OrigCost -= DupCost; 3321 if (OrigCost > BBDupThreshold) { 3322 Candidates.push_back(Pred); 3323 if (SuccIt != Succs.end()) 3324 SuccIt++; 3325 } 3326 } 3327 3328 // No predecessors can optimally fallthrough to BB. 3329 // So we can change one duplication into fallthrough. 3330 if (!Fallthrough) { 3331 if ((Candidates.size() < Preds.size()) && (Candidates.size() > 0)) { 3332 Candidates[0] = Candidates.back(); 3333 Candidates.pop_back(); 3334 } 3335 } 3336 } 3337 3338 void MachineBlockPlacement::initDupThreshold() { 3339 DupThreshold = BlockFrequency(0); 3340 if (!F->getFunction().hasProfileData()) 3341 return; 3342 3343 // We prefer to use prifile count. 3344 uint64_t HotThreshold = PSI->getOrCompHotCountThreshold(); 3345 if (HotThreshold != UINT64_MAX) { 3346 UseProfileCount = true; 3347 DupThreshold = 3348 BlockFrequency(HotThreshold * TailDupProfilePercentThreshold / 100); 3349 return; 3350 } 3351 3352 // Profile count is not available, we can use block frequency instead. 3353 BlockFrequency MaxFreq = BlockFrequency(0); 3354 for (MachineBasicBlock &MBB : *F) { 3355 BlockFrequency Freq = MBFI->getBlockFreq(&MBB); 3356 if (Freq > MaxFreq) 3357 MaxFreq = Freq; 3358 } 3359 3360 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100); 3361 DupThreshold = BlockFrequency(MaxFreq * ThresholdProb); 3362 UseProfileCount = false; 3363 } 3364 3365 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) { 3366 if (skipFunction(MF.getFunction())) 3367 return false; 3368 3369 // Check for single-block functions and skip them. 3370 if (std::next(MF.begin()) == MF.end()) 3371 return false; 3372 3373 F = &MF; 3374 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 3375 MBFI = std::make_unique<MBFIWrapper>( 3376 getAnalysis<MachineBlockFrequencyInfo>()); 3377 MLI = &getAnalysis<MachineLoopInfo>(); 3378 TII = MF.getSubtarget().getInstrInfo(); 3379 TLI = MF.getSubtarget().getTargetLowering(); 3380 MPDT = nullptr; 3381 PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 3382 3383 initDupThreshold(); 3384 3385 // Initialize PreferredLoopExit to nullptr here since it may never be set if 3386 // there are no MachineLoops. 3387 PreferredLoopExit = nullptr; 3388 3389 assert(BlockToChain.empty() && 3390 "BlockToChain map should be empty before starting placement."); 3391 assert(ComputedEdges.empty() && 3392 "Computed Edge map should be empty before starting placement."); 3393 3394 unsigned TailDupSize = TailDupPlacementThreshold; 3395 // If only the aggressive threshold is explicitly set, use it. 3396 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 && 3397 TailDupPlacementThreshold.getNumOccurrences() == 0) 3398 TailDupSize = TailDupPlacementAggressiveThreshold; 3399 3400 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>(); 3401 // For aggressive optimization, we can adjust some thresholds to be less 3402 // conservative. 3403 if (PassConfig->getOptLevel() >= CodeGenOptLevel::Aggressive) { 3404 // At O3 we should be more willing to copy blocks for tail duplication. This 3405 // increases size pressure, so we only do it at O3 3406 // Do this unless only the regular threshold is explicitly set. 3407 if (TailDupPlacementThreshold.getNumOccurrences() == 0 || 3408 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0) 3409 TailDupSize = TailDupPlacementAggressiveThreshold; 3410 } 3411 3412 // If there's no threshold provided through options, query the target 3413 // information for a threshold instead. 3414 if (TailDupPlacementThreshold.getNumOccurrences() == 0 && 3415 (PassConfig->getOptLevel() < CodeGenOptLevel::Aggressive || 3416 TailDupPlacementAggressiveThreshold.getNumOccurrences() == 0)) 3417 TailDupSize = TII->getTailDuplicateSize(PassConfig->getOptLevel()); 3418 3419 if (allowTailDupPlacement()) { 3420 MPDT = &getAnalysis<MachinePostDominatorTree>(); 3421 bool OptForSize = MF.getFunction().hasOptSize() || 3422 llvm::shouldOptimizeForSize(&MF, PSI, &MBFI->getMBFI()); 3423 if (OptForSize) 3424 TailDupSize = 1; 3425 bool PreRegAlloc = false; 3426 TailDup.initMF(MF, PreRegAlloc, MBPI, MBFI.get(), PSI, 3427 /* LayoutMode */ true, TailDupSize); 3428 precomputeTriangleChains(); 3429 } 3430 3431 buildCFGChains(); 3432 3433 // Changing the layout can create new tail merging opportunities. 3434 // TailMerge can create jump into if branches that make CFG irreducible for 3435 // HW that requires structured CFG. 3436 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() && 3437 PassConfig->getEnableTailMerge() && 3438 BranchFoldPlacement; 3439 // No tail merging opportunities if the block number is less than four. 3440 if (MF.size() > 3 && EnableTailMerge) { 3441 unsigned TailMergeSize = TailDupSize + 1; 3442 BranchFolder BF(/*DefaultEnableTailMerge=*/true, /*CommonHoist=*/false, 3443 *MBFI, *MBPI, PSI, TailMergeSize); 3444 3445 if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(), MLI, 3446 /*AfterPlacement=*/true)) { 3447 // Redo the layout if tail merging creates/removes/moves blocks. 3448 BlockToChain.clear(); 3449 ComputedEdges.clear(); 3450 // Must redo the post-dominator tree if blocks were changed. 3451 if (MPDT) 3452 MPDT->runOnMachineFunction(MF); 3453 ChainAllocator.DestroyAll(); 3454 buildCFGChains(); 3455 } 3456 } 3457 3458 // Apply a post-processing optimizing block placement. 3459 if (MF.size() >= 3 && EnableExtTspBlockPlacement && 3460 (ApplyExtTspWithoutProfile || MF.getFunction().hasProfileData())) { 3461 // Find a new placement and modify the layout of the blocks in the function. 3462 applyExtTsp(); 3463 3464 // Re-create CFG chain so that we can optimizeBranches and alignBlocks. 3465 createCFGChainExtTsp(); 3466 } 3467 3468 optimizeBranches(); 3469 alignBlocks(); 3470 3471 BlockToChain.clear(); 3472 ComputedEdges.clear(); 3473 ChainAllocator.DestroyAll(); 3474 3475 bool HasMaxBytesOverride = 3476 MaxBytesForAlignmentOverride.getNumOccurrences() > 0; 3477 3478 if (AlignAllBlock) 3479 // Align all of the blocks in the function to a specific alignment. 3480 for (MachineBasicBlock &MBB : MF) { 3481 if (HasMaxBytesOverride) 3482 MBB.setAlignment(Align(1ULL << AlignAllBlock), 3483 MaxBytesForAlignmentOverride); 3484 else 3485 MBB.setAlignment(Align(1ULL << AlignAllBlock)); 3486 } 3487 else if (AlignAllNonFallThruBlocks) { 3488 // Align all of the blocks that have no fall-through predecessors to a 3489 // specific alignment. 3490 for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) { 3491 auto LayoutPred = std::prev(MBI); 3492 if (!LayoutPred->isSuccessor(&*MBI)) { 3493 if (HasMaxBytesOverride) 3494 MBI->setAlignment(Align(1ULL << AlignAllNonFallThruBlocks), 3495 MaxBytesForAlignmentOverride); 3496 else 3497 MBI->setAlignment(Align(1ULL << AlignAllNonFallThruBlocks)); 3498 } 3499 } 3500 } 3501 if (ViewBlockLayoutWithBFI != GVDT_None && 3502 (ViewBlockFreqFuncName.empty() || 3503 F->getFunction().getName().equals(ViewBlockFreqFuncName))) { 3504 if (RenumberBlocksBeforeView) 3505 MF.RenumberBlocks(); 3506 MBFI->view("MBP." + MF.getName(), false); 3507 } 3508 3509 // We always return true as we have no way to track whether the final order 3510 // differs from the original order. 3511 return true; 3512 } 3513 3514 void MachineBlockPlacement::applyExtTsp() { 3515 // Prepare data; blocks are indexed by their index in the current ordering. 3516 DenseMap<const MachineBasicBlock *, uint64_t> BlockIndex; 3517 BlockIndex.reserve(F->size()); 3518 std::vector<const MachineBasicBlock *> CurrentBlockOrder; 3519 CurrentBlockOrder.reserve(F->size()); 3520 size_t NumBlocks = 0; 3521 for (const MachineBasicBlock &MBB : *F) { 3522 BlockIndex[&MBB] = NumBlocks++; 3523 CurrentBlockOrder.push_back(&MBB); 3524 } 3525 3526 auto BlockSizes = std::vector<uint64_t>(F->size()); 3527 auto BlockCounts = std::vector<uint64_t>(F->size()); 3528 std::vector<codelayout::EdgeCount> JumpCounts; 3529 for (MachineBasicBlock &MBB : *F) { 3530 // Getting the block frequency. 3531 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB); 3532 BlockCounts[BlockIndex[&MBB]] = BlockFreq.getFrequency(); 3533 // Getting the block size: 3534 // - approximate the size of an instruction by 4 bytes, and 3535 // - ignore debug instructions. 3536 // Note: getting the exact size of each block is target-dependent and can be 3537 // done by extending the interface of MCCodeEmitter. Experimentally we do 3538 // not see a perf improvement with the exact block sizes. 3539 auto NonDbgInsts = 3540 instructionsWithoutDebug(MBB.instr_begin(), MBB.instr_end()); 3541 int NumInsts = std::distance(NonDbgInsts.begin(), NonDbgInsts.end()); 3542 BlockSizes[BlockIndex[&MBB]] = 4 * NumInsts; 3543 // Getting jump frequencies. 3544 for (MachineBasicBlock *Succ : MBB.successors()) { 3545 auto EP = MBPI->getEdgeProbability(&MBB, Succ); 3546 BlockFrequency JumpFreq = BlockFreq * EP; 3547 JumpCounts.push_back( 3548 {BlockIndex[&MBB], BlockIndex[Succ], JumpFreq.getFrequency()}); 3549 } 3550 } 3551 3552 LLVM_DEBUG(dbgs() << "Applying ext-tsp layout for |V| = " << F->size() 3553 << " with profile = " << F->getFunction().hasProfileData() 3554 << " (" << F->getName().str() << ")" 3555 << "\n"); 3556 LLVM_DEBUG( 3557 dbgs() << format(" original layout score: %0.2f\n", 3558 calcExtTspScore(BlockSizes, BlockCounts, JumpCounts))); 3559 3560 // Run the layout algorithm. 3561 auto NewOrder = computeExtTspLayout(BlockSizes, BlockCounts, JumpCounts); 3562 std::vector<const MachineBasicBlock *> NewBlockOrder; 3563 NewBlockOrder.reserve(F->size()); 3564 for (uint64_t Node : NewOrder) { 3565 NewBlockOrder.push_back(CurrentBlockOrder[Node]); 3566 } 3567 LLVM_DEBUG(dbgs() << format(" optimized layout score: %0.2f\n", 3568 calcExtTspScore(NewOrder, BlockSizes, BlockCounts, 3569 JumpCounts))); 3570 3571 // Assign new block order. 3572 assignBlockOrder(NewBlockOrder); 3573 } 3574 3575 void MachineBlockPlacement::assignBlockOrder( 3576 const std::vector<const MachineBasicBlock *> &NewBlockOrder) { 3577 assert(F->size() == NewBlockOrder.size() && "Incorrect size of block order"); 3578 F->RenumberBlocks(); 3579 3580 bool HasChanges = false; 3581 for (size_t I = 0; I < NewBlockOrder.size(); I++) { 3582 if (NewBlockOrder[I] != F->getBlockNumbered(I)) { 3583 HasChanges = true; 3584 break; 3585 } 3586 } 3587 // Stop early if the new block order is identical to the existing one. 3588 if (!HasChanges) 3589 return; 3590 3591 SmallVector<MachineBasicBlock *, 4> PrevFallThroughs(F->getNumBlockIDs()); 3592 for (auto &MBB : *F) { 3593 PrevFallThroughs[MBB.getNumber()] = MBB.getFallThrough(); 3594 } 3595 3596 // Sort basic blocks in the function according to the computed order. 3597 DenseMap<const MachineBasicBlock *, size_t> NewIndex; 3598 for (const MachineBasicBlock *MBB : NewBlockOrder) { 3599 NewIndex[MBB] = NewIndex.size(); 3600 } 3601 F->sort([&](MachineBasicBlock &L, MachineBasicBlock &R) { 3602 return NewIndex[&L] < NewIndex[&R]; 3603 }); 3604 3605 // Update basic block branches by inserting explicit fallthrough branches 3606 // when required and re-optimize branches when possible. 3607 const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo(); 3608 SmallVector<MachineOperand, 4> Cond; 3609 for (auto &MBB : *F) { 3610 MachineFunction::iterator NextMBB = std::next(MBB.getIterator()); 3611 MachineFunction::iterator EndIt = MBB.getParent()->end(); 3612 auto *FTMBB = PrevFallThroughs[MBB.getNumber()]; 3613 // If this block had a fallthrough before we need an explicit unconditional 3614 // branch to that block if the fallthrough block is not adjacent to the 3615 // block in the new order. 3616 if (FTMBB && (NextMBB == EndIt || &*NextMBB != FTMBB)) { 3617 TII->insertUnconditionalBranch(MBB, FTMBB, MBB.findBranchDebugLoc()); 3618 } 3619 3620 // It might be possible to optimize branches by flipping the condition. 3621 Cond.clear(); 3622 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 3623 if (TII->analyzeBranch(MBB, TBB, FBB, Cond)) 3624 continue; 3625 MBB.updateTerminator(FTMBB); 3626 } 3627 3628 #ifndef NDEBUG 3629 // Make sure we correctly constructed all branches. 3630 F->verify(this, "After optimized block reordering"); 3631 #endif 3632 } 3633 3634 void MachineBlockPlacement::createCFGChainExtTsp() { 3635 BlockToChain.clear(); 3636 ComputedEdges.clear(); 3637 ChainAllocator.DestroyAll(); 3638 3639 MachineBasicBlock *HeadBB = &F->front(); 3640 BlockChain *FunctionChain = 3641 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, HeadBB); 3642 3643 for (MachineBasicBlock &MBB : *F) { 3644 if (HeadBB == &MBB) 3645 continue; // Ignore head of the chain 3646 FunctionChain->merge(&MBB, nullptr); 3647 } 3648 } 3649 3650 namespace { 3651 3652 /// A pass to compute block placement statistics. 3653 /// 3654 /// A separate pass to compute interesting statistics for evaluating block 3655 /// placement. This is separate from the actual placement pass so that they can 3656 /// be computed in the absence of any placement transformations or when using 3657 /// alternative placement strategies. 3658 class MachineBlockPlacementStats : public MachineFunctionPass { 3659 /// A handle to the branch probability pass. 3660 const MachineBranchProbabilityInfo *MBPI; 3661 3662 /// A handle to the function-wide block frequency pass. 3663 const MachineBlockFrequencyInfo *MBFI; 3664 3665 public: 3666 static char ID; // Pass identification, replacement for typeid 3667 3668 MachineBlockPlacementStats() : MachineFunctionPass(ID) { 3669 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry()); 3670 } 3671 3672 bool runOnMachineFunction(MachineFunction &F) override; 3673 3674 void getAnalysisUsage(AnalysisUsage &AU) const override { 3675 AU.addRequired<MachineBranchProbabilityInfo>(); 3676 AU.addRequired<MachineBlockFrequencyInfo>(); 3677 AU.setPreservesAll(); 3678 MachineFunctionPass::getAnalysisUsage(AU); 3679 } 3680 }; 3681 3682 } // end anonymous namespace 3683 3684 char MachineBlockPlacementStats::ID = 0; 3685 3686 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID; 3687 3688 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats", 3689 "Basic Block Placement Stats", false, false) 3690 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 3691 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 3692 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats", 3693 "Basic Block Placement Stats", false, false) 3694 3695 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) { 3696 // Check for single-block functions and skip them. 3697 if (std::next(F.begin()) == F.end()) 3698 return false; 3699 3700 if (!isFunctionInPrintList(F.getName())) 3701 return false; 3702 3703 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 3704 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 3705 3706 for (MachineBasicBlock &MBB : F) { 3707 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB); 3708 Statistic &NumBranches = 3709 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches; 3710 Statistic &BranchTakenFreq = 3711 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq; 3712 for (MachineBasicBlock *Succ : MBB.successors()) { 3713 // Skip if this successor is a fallthrough. 3714 if (MBB.isLayoutSuccessor(Succ)) 3715 continue; 3716 3717 BlockFrequency EdgeFreq = 3718 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ); 3719 ++NumBranches; 3720 BranchTakenFreq += EdgeFreq.getFrequency(); 3721 } 3722 } 3723 3724 return false; 3725 } 3726