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