MachineBlockPlacement.cpp revision fa97658b1c71f747cfe0f3e1f1bcbd86d7fa9f75
1//===-- MachineBlockPlacement.cpp - Basic Block Code Layout optimization --===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements basic block placement transformations using the CFG 11// structure and branch probability estimates. 12// 13// The pass strives to preserve the structure of the CFG (that is, retain 14// a topological ordering of basic blocks) in the absense of a *strong* signal 15// to the contrary from probabilities. However, within the CFG structure, it 16// attempts to choose an ordering which favors placing more likely sequences of 17// blocks adjacent to each other. 18// 19// The algorithm works from the inner-most loop within a function outward, and 20// at each stage walks through the basic blocks, trying to coalesce them into 21// sequential chains where allowed by the CFG (or demanded by heavy 22// probabilities). Finally, it walks the blocks in topological order, and the 23// first time it reaches a chain of basic blocks, it schedules them in the 24// function in-order. 25// 26//===----------------------------------------------------------------------===// 27 28#define DEBUG_TYPE "block-placement2" 29#include "llvm/CodeGen/MachineBasicBlock.h" 30#include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 31#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 32#include "llvm/CodeGen/MachineFunction.h" 33#include "llvm/CodeGen/MachineFunctionPass.h" 34#include "llvm/CodeGen/MachineLoopInfo.h" 35#include "llvm/CodeGen/MachineModuleInfo.h" 36#include "llvm/CodeGen/Passes.h" 37#include "llvm/Support/Allocator.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/ErrorHandling.h" 40#include "llvm/ADT/DenseMap.h" 41#include "llvm/ADT/PostOrderIterator.h" 42#include "llvm/ADT/SCCIterator.h" 43#include "llvm/ADT/SmallPtrSet.h" 44#include "llvm/ADT/SmallVector.h" 45#include "llvm/ADT/Statistic.h" 46#include "llvm/Target/TargetInstrInfo.h" 47#include "llvm/Target/TargetLowering.h" 48#include <algorithm> 49using namespace llvm; 50 51STATISTIC(NumCondBranches, "Number of conditional branches"); 52STATISTIC(NumUncondBranches, "Number of uncondittional branches"); 53STATISTIC(CondBranchTakenFreq, 54 "Potential frequency of taking conditional branches"); 55STATISTIC(UncondBranchTakenFreq, 56 "Potential frequency of taking unconditional branches"); 57 58namespace { 59/// \brief A structure for storing a weighted edge. 60/// 61/// This stores an edge and its weight, computed as the product of the 62/// frequency that the starting block is entered with the probability of 63/// a particular exit block. 64struct WeightedEdge { 65 BlockFrequency EdgeFrequency; 66 MachineBasicBlock *From, *To; 67 68 bool operator<(const WeightedEdge &RHS) const { 69 return EdgeFrequency < RHS.EdgeFrequency; 70 } 71}; 72} 73 74namespace { 75class BlockChain; 76/// \brief Type for our function-wide basic block -> block chain mapping. 77typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType; 78} 79 80namespace { 81/// \brief A chain of blocks which will be laid out contiguously. 82/// 83/// This is the datastructure representing a chain of consecutive blocks that 84/// are profitable to layout together in order to maximize fallthrough 85/// probabilities. We also can use a block chain to represent a sequence of 86/// basic blocks which have some external (correctness) requirement for 87/// sequential layout. 88/// 89/// Eventually, the block chains will form a directed graph over the function. 90/// We provide an SCC-supporting-iterator in order to quicky build and walk the 91/// SCCs of block chains within a function. 92/// 93/// The block chains also have support for calculating and caching probability 94/// information related to the chain itself versus other chains. This is used 95/// for ranking during the final layout of block chains. 96class BlockChain { 97 /// \brief The sequence of blocks belonging to this chain. 98 /// 99 /// This is the sequence of blocks for a particular chain. These will be laid 100 /// out in-order within the function. 101 SmallVector<MachineBasicBlock *, 4> Blocks; 102 103 /// \brief A handle to the function-wide basic block to block chain mapping. 104 /// 105 /// This is retained in each block chain to simplify the computation of child 106 /// block chains for SCC-formation and iteration. We store the edges to child 107 /// basic blocks, and map them back to their associated chains using this 108 /// structure. 109 BlockToChainMapType &BlockToChain; 110 111public: 112 /// \brief Construct a new BlockChain. 113 /// 114 /// This builds a new block chain representing a single basic block in the 115 /// function. It also registers itself as the chain that block participates 116 /// in with the BlockToChain mapping. 117 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB) 118 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) { 119 assert(BB && "Cannot create a chain with a null basic block"); 120 BlockToChain[BB] = this; 121 } 122 123 /// \brief Iterator over blocks within the chain. 124 typedef SmallVectorImpl<MachineBasicBlock *>::const_iterator iterator; 125 126 /// \brief Beginning of blocks within the chain. 127 iterator begin() const { return Blocks.begin(); } 128 129 /// \brief End of blocks within the chain. 130 iterator end() const { return Blocks.end(); } 131 132 /// \brief Merge a block chain into this one. 133 /// 134 /// This routine merges a block chain into this one. It takes care of forming 135 /// a contiguous sequence of basic blocks, updating the edge list, and 136 /// updating the block -> chain mapping. It does not free or tear down the 137 /// old chain, but the old chain's block list is no longer valid. 138 void merge(MachineBasicBlock *BB, BlockChain *Chain) { 139 assert(BB); 140 assert(!Blocks.empty()); 141 142 // Fast path in case we don't have a chain already. 143 if (!Chain) { 144 assert(!BlockToChain[BB]); 145 Blocks.push_back(BB); 146 BlockToChain[BB] = this; 147 return; 148 } 149 150 assert(BB == *Chain->begin()); 151 assert(Chain->begin() != Chain->end()); 152 153 // Update the incoming blocks to point to this chain, and add them to the 154 // chain structure. 155 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end(); 156 BI != BE; ++BI) { 157 Blocks.push_back(*BI); 158 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain"); 159 BlockToChain[*BI] = this; 160 } 161 } 162 163 /// \brief Count of predecessors within the loop currently being processed. 164 /// 165 /// This count is updated at each loop we process to represent the number of 166 /// in-loop predecessors of this chain. 167 unsigned LoopPredecessors; 168}; 169} 170 171namespace { 172class MachineBlockPlacement : public MachineFunctionPass { 173 /// \brief A typedef for a block filter set. 174 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet; 175 176 /// \brief A handle to the branch probability pass. 177 const MachineBranchProbabilityInfo *MBPI; 178 179 /// \brief A handle to the function-wide block frequency pass. 180 const MachineBlockFrequencyInfo *MBFI; 181 182 /// \brief A handle to the loop info. 183 const MachineLoopInfo *MLI; 184 185 /// \brief A handle to the target's instruction info. 186 const TargetInstrInfo *TII; 187 188 /// \brief A handle to the target's lowering info. 189 const TargetLowering *TLI; 190 191 /// \brief Allocator and owner of BlockChain structures. 192 /// 193 /// We build BlockChains lazily by merging together high probability BB 194 /// sequences acording to the "Algo2" in the paper mentioned at the top of 195 /// the file. To reduce malloc traffic, we allocate them using this slab-like 196 /// allocator, and destroy them after the pass completes. 197 SpecificBumpPtrAllocator<BlockChain> ChainAllocator; 198 199 /// \brief Function wide BasicBlock to BlockChain mapping. 200 /// 201 /// This mapping allows efficiently moving from any given basic block to the 202 /// BlockChain it participates in, if any. We use it to, among other things, 203 /// allow implicitly defining edges between chains as the existing edges 204 /// between basic blocks. 205 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain; 206 207 void markChainSuccessors(BlockChain &Chain, 208 MachineBasicBlock *LoopHeaderBB, 209 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 210 const BlockFilterSet *BlockFilter = 0); 211 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB, 212 BlockChain &Chain, 213 const BlockFilterSet *BlockFilter); 214 MachineBasicBlock *selectBestCandidateBlock( 215 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList, 216 const BlockFilterSet *BlockFilter); 217 MachineBasicBlock *getFirstUnplacedBlock(const BlockChain &PlacedChain, 218 ArrayRef<MachineBasicBlock *> Blocks, 219 unsigned &PrevUnplacedBlockIdx); 220 void buildChain(MachineBasicBlock *BB, BlockChain &Chain, 221 ArrayRef<MachineBasicBlock *> Blocks, 222 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 223 const BlockFilterSet *BlockFilter = 0); 224 void buildLoopChains(MachineFunction &F, MachineLoop &L); 225 void buildCFGChains(MachineFunction &F); 226 void AlignLoops(MachineFunction &F); 227 228public: 229 static char ID; // Pass identification, replacement for typeid 230 MachineBlockPlacement() : MachineFunctionPass(ID) { 231 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry()); 232 } 233 234 bool runOnMachineFunction(MachineFunction &F); 235 236 void getAnalysisUsage(AnalysisUsage &AU) const { 237 AU.addRequired<MachineBranchProbabilityInfo>(); 238 AU.addRequired<MachineBlockFrequencyInfo>(); 239 AU.addRequired<MachineLoopInfo>(); 240 MachineFunctionPass::getAnalysisUsage(AU); 241 } 242 243 const char *getPassName() const { return "Block Placement"; } 244}; 245} 246 247char MachineBlockPlacement::ID = 0; 248INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2", 249 "Branch Probability Basic Block Placement", false, false) 250INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 251INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 252INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 253INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2", 254 "Branch Probability Basic Block Placement", false, false) 255 256FunctionPass *llvm::createMachineBlockPlacementPass() { 257 return new MachineBlockPlacement(); 258} 259 260#ifndef NDEBUG 261/// \brief Helper to print the name of a MBB. 262/// 263/// Only used by debug logging. 264static std::string getBlockName(MachineBasicBlock *BB) { 265 std::string Result; 266 raw_string_ostream OS(Result); 267 OS << "BB#" << BB->getNumber() 268 << " (derived from LLVM BB '" << BB->getName() << "')"; 269 OS.flush(); 270 return Result; 271} 272 273/// \brief Helper to print the number of a MBB. 274/// 275/// Only used by debug logging. 276static std::string getBlockNum(MachineBasicBlock *BB) { 277 std::string Result; 278 raw_string_ostream OS(Result); 279 OS << "BB#" << BB->getNumber(); 280 OS.flush(); 281 return Result; 282} 283#endif 284 285/// \brief Mark a chain's successors as having one fewer preds. 286/// 287/// When a chain is being merged into the "placed" chain, this routine will 288/// quickly walk the successors of each block in the chain and mark them as 289/// having one fewer active predecessor. It also adds any successors of this 290/// chain which reach the zero-predecessor state to the worklist passed in. 291void MachineBlockPlacement::markChainSuccessors( 292 BlockChain &Chain, 293 MachineBasicBlock *LoopHeaderBB, 294 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 295 const BlockFilterSet *BlockFilter) { 296 // Walk all the blocks in this chain, marking their successors as having 297 // a predecessor placed. 298 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end(); 299 CBI != CBE; ++CBI) { 300 // Add any successors for which this is the only un-placed in-loop 301 // predecessor to the worklist as a viable candidate for CFG-neutral 302 // placement. No subsequent placement of this block will violate the CFG 303 // shape, so we get to use heuristics to choose a favorable placement. 304 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(), 305 SE = (*CBI)->succ_end(); 306 SI != SE; ++SI) { 307 if (BlockFilter && !BlockFilter->count(*SI)) 308 continue; 309 BlockChain &SuccChain = *BlockToChain[*SI]; 310 // Disregard edges within a fixed chain, or edges to the loop header. 311 if (&Chain == &SuccChain || *SI == LoopHeaderBB) 312 continue; 313 314 // This is a cross-chain edge that is within the loop, so decrement the 315 // loop predecessor count of the destination chain. 316 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0) 317 BlockWorkList.push_back(*SI); 318 } 319 } 320} 321 322/// \brief Select the best successor for a block. 323/// 324/// This looks across all successors of a particular block and attempts to 325/// select the "best" one to be the layout successor. It only considers direct 326/// successors which also pass the block filter. It will attempt to avoid 327/// breaking CFG structure, but cave and break such structures in the case of 328/// very hot successor edges. 329/// 330/// \returns The best successor block found, or null if none are viable. 331MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor( 332 MachineBasicBlock *BB, BlockChain &Chain, 333 const BlockFilterSet *BlockFilter) { 334 const BranchProbability HotProb(4, 5); // 80% 335 336 MachineBasicBlock *BestSucc = 0; 337 // FIXME: Due to the performance of the probability and weight routines in 338 // the MBPI analysis, we manually compute probabilities using the edge 339 // weights. This is suboptimal as it means that the somewhat subtle 340 // definition of edge weight semantics is encoded here as well. We should 341 // improve the MBPI interface to effeciently support query patterns such as 342 // this. 343 uint32_t BestWeight = 0; 344 uint32_t WeightScale = 0; 345 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale); 346 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n"); 347 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 348 SE = BB->succ_end(); 349 SI != SE; ++SI) { 350 if (BlockFilter && !BlockFilter->count(*SI)) 351 continue; 352 BlockChain &SuccChain = *BlockToChain[*SI]; 353 if (&SuccChain == &Chain) { 354 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n"); 355 continue; 356 } 357 358 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI); 359 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight); 360 361 // Only consider successors which are either "hot", or wouldn't violate 362 // any CFG constraints. 363 if (SuccChain.LoopPredecessors != 0 && SuccProb < HotProb) { 364 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n"); 365 continue; 366 } 367 368 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb 369 << " (prob)" 370 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "") 371 << "\n"); 372 if (BestSucc && BestWeight >= SuccWeight) 373 continue; 374 BestSucc = *SI; 375 BestWeight = SuccWeight; 376 } 377 return BestSucc; 378} 379 380namespace { 381/// \brief Predicate struct to detect blocks already placed. 382class IsBlockPlaced { 383 const BlockChain &PlacedChain; 384 const BlockToChainMapType &BlockToChain; 385 386public: 387 IsBlockPlaced(const BlockChain &PlacedChain, 388 const BlockToChainMapType &BlockToChain) 389 : PlacedChain(PlacedChain), BlockToChain(BlockToChain) {} 390 391 bool operator()(MachineBasicBlock *BB) const { 392 return BlockToChain.lookup(BB) == &PlacedChain; 393 } 394}; 395} 396 397/// \brief Select the best block from a worklist. 398/// 399/// This looks through the provided worklist as a list of candidate basic 400/// blocks and select the most profitable one to place. The definition of 401/// profitable only really makes sense in the context of a loop. This returns 402/// the most frequently visited block in the worklist, which in the case of 403/// a loop, is the one most desirable to be physically close to the rest of the 404/// loop body in order to improve icache behavior. 405/// 406/// \returns The best block found, or null if none are viable. 407MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock( 408 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList, 409 const BlockFilterSet *BlockFilter) { 410 // Once we need to walk the worklist looking for a candidate, cleanup the 411 // worklist of already placed entries. 412 // FIXME: If this shows up on profiles, it could be folded (at the cost of 413 // some code complexity) into the loop below. 414 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(), 415 IsBlockPlaced(Chain, BlockToChain)), 416 WorkList.end()); 417 418 MachineBasicBlock *BestBlock = 0; 419 BlockFrequency BestFreq; 420 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(), 421 WBE = WorkList.end(); 422 WBI != WBE; ++WBI) { 423 assert(!BlockFilter || BlockFilter->count(*WBI)); 424 BlockChain &SuccChain = *BlockToChain[*WBI]; 425 if (&SuccChain == &Chain) { 426 DEBUG(dbgs() << " " << getBlockName(*WBI) 427 << " -> Already merged!\n"); 428 continue; 429 } 430 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block"); 431 432 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI); 433 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq 434 << " (freq)\n"); 435 if (BestBlock && BestFreq >= CandidateFreq) 436 continue; 437 BestBlock = *WBI; 438 BestFreq = CandidateFreq; 439 } 440 return BestBlock; 441} 442 443/// \brief Retrieve the first unplaced basic block. 444/// 445/// This routine is called when we are unable to use the CFG to walk through 446/// all of the basic blocks and form a chain due to unnatural loops in the CFG. 447/// We walk through the sequence of blocks, starting from the 448/// LastUnplacedBlockIdx. We update this index to avoid re-scanning the entire 449/// sequence on repeated calls to this routine. 450MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock( 451 const BlockChain &PlacedChain, 452 ArrayRef<MachineBasicBlock *> Blocks, 453 unsigned &PrevUnplacedBlockIdx) { 454 for (unsigned i = PrevUnplacedBlockIdx, e = Blocks.size(); i != e; ++i) { 455 MachineBasicBlock *BB = Blocks[i]; 456 if (BlockToChain[BB] != &PlacedChain) { 457 PrevUnplacedBlockIdx = i; 458 return BB; 459 } 460 } 461 return 0; 462} 463 464void MachineBlockPlacement::buildChain( 465 MachineBasicBlock *BB, 466 BlockChain &Chain, 467 ArrayRef<MachineBasicBlock *> Blocks, 468 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 469 const BlockFilterSet *BlockFilter) { 470 assert(BB); 471 assert(BlockToChain[BB] == &Chain); 472 assert(*Chain.begin() == BB); 473 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch. 474 unsigned PrevUnplacedBlockIdx = 0; 475 476 MachineBasicBlock *LoopHeaderBB = BB; 477 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter); 478 BB = *llvm::prior(Chain.end()); 479 for (;;) { 480 assert(BB); 481 assert(BlockToChain[BB] == &Chain); 482 assert(*llvm::prior(Chain.end()) == BB); 483 MachineBasicBlock *BestSucc = 0; 484 485 // Check for unreasonable branches, and forcibly merge the existing layout 486 // successor for them. We can handle cases that AnalyzeBranch can't: jump 487 // tables etc are fine. The case we want to handle specially is when there 488 // is potential fallthrough, but the branch cannot be analyzed. This 489 // includes blocks without terminators as well as other cases. 490 Cond.clear(); 491 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 492 if (TII->AnalyzeBranch(*BB, TBB, FBB, Cond) && BB->canFallThrough()) { 493 MachineFunction::iterator I(BB), NextI(llvm::next(I)); 494 // Ensure that the layout successor is a viable block, as we know that 495 // fallthrough is a possibility. 496 assert(NextI != BB->getParent()->end()); 497 assert(!BlockFilter || BlockFilter->count(NextI)); 498 BestSucc = NextI; 499 } 500 501 // Otherwise, look for the best viable successor if there is one to place 502 // immediately after this block. 503 if (!BestSucc) 504 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter); 505 506 // If an immediate successor isn't available, look for the best viable 507 // block among those we've identified as not violating the loop's CFG at 508 // this point. This won't be a fallthrough, but it will increase locality. 509 if (!BestSucc) 510 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter); 511 512 if (!BestSucc) { 513 BestSucc = getFirstUnplacedBlock(Chain, Blocks, PrevUnplacedBlockIdx); 514 if (!BestSucc) 515 break; 516 517 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the " 518 "layout successor until the CFG reduces\n"); 519 } 520 521 // Place this block, updating the datastructures to reflect its placement. 522 BlockChain &SuccChain = *BlockToChain[BestSucc]; 523 // Zero out LoopPredecessors for the successor we're about to merge in case 524 // we selected a successor that didn't fit naturally into the CFG. 525 SuccChain.LoopPredecessors = 0; 526 DEBUG(dbgs() << "Merging from " << getBlockNum(BB) 527 << " to " << getBlockNum(BestSucc) << "\n"); 528 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter); 529 Chain.merge(BestSucc, &SuccChain); 530 BB = *llvm::prior(Chain.end()); 531 }; 532 533 DEBUG(dbgs() << "Finished forming chain for header block " 534 << getBlockNum(*Chain.begin()) << "\n"); 535} 536 537/// \brief Forms basic block chains from the natural loop structures. 538/// 539/// These chains are designed to preserve the existing *structure* of the code 540/// as much as possible. We can then stitch the chains together in a way which 541/// both preserves the topological structure and minimizes taken conditional 542/// branches. 543void MachineBlockPlacement::buildLoopChains(MachineFunction &F, 544 MachineLoop &L) { 545 // First recurse through any nested loops, building chains for those inner 546 // loops. 547 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI) 548 buildLoopChains(F, **LI); 549 550 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 551 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end()); 552 BlockChain &LoopChain = *BlockToChain[L.getHeader()]; 553 554 // FIXME: This is a really lame way of walking the chains in the loop: we 555 // walk the blocks, and use a set to prevent visiting a particular chain 556 // twice. 557 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 558 for (MachineLoop::block_iterator BI = L.block_begin(), 559 BE = L.block_end(); 560 BI != BE; ++BI) { 561 BlockChain &Chain = *BlockToChain[*BI]; 562 if (!UpdatedPreds.insert(&Chain) || BI == L.block_begin()) 563 continue; 564 565 assert(Chain.LoopPredecessors == 0); 566 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end(); 567 BCI != BCE; ++BCI) { 568 assert(BlockToChain[*BCI] == &Chain); 569 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(), 570 PE = (*BCI)->pred_end(); 571 PI != PE; ++PI) { 572 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI)) 573 continue; 574 ++Chain.LoopPredecessors; 575 } 576 } 577 578 if (Chain.LoopPredecessors == 0) 579 BlockWorkList.push_back(*BI); 580 } 581 582 buildChain(*L.block_begin(), LoopChain, L.getBlocks(), BlockWorkList, 583 &LoopBlockSet); 584 585 DEBUG({ 586 // Crash at the end so we get all of the debugging output first. 587 bool BadLoop = false; 588 if (LoopChain.LoopPredecessors) { 589 BadLoop = true; 590 dbgs() << "Loop chain contains a block without its preds placed!\n" 591 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 592 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"; 593 } 594 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end(); 595 BCI != BCE; ++BCI) 596 if (!LoopBlockSet.erase(*BCI)) { 597 BadLoop = true; 598 dbgs() << "Loop chain contains a block not contained by the loop!\n" 599 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 600 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 601 << " Bad block: " << getBlockName(*BCI) << "\n"; 602 } 603 604 if (!LoopBlockSet.empty()) { 605 BadLoop = true; 606 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(), 607 LBE = LoopBlockSet.end(); 608 LBI != LBE; ++LBI) 609 dbgs() << "Loop contains blocks never placed into a chain!\n" 610 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 611 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 612 << " Bad block: " << getBlockName(*LBI) << "\n"; 613 } 614 assert(!BadLoop && "Detected problems with the placement of this loop."); 615 }); 616} 617 618void MachineBlockPlacement::buildCFGChains(MachineFunction &F) { 619 // Ensure that every BB in the function has an associated chain to simplify 620 // the assumptions of the remaining algorithm. 621 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 622 BlockToChain[&*FI] = 623 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, &*FI); 624 625 // Build any loop-based chains. 626 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE; 627 ++LI) 628 buildLoopChains(F, **LI); 629 630 // We need a vector of blocks so that buildChain can handle unnatural CFG 631 // constructs by searching for unplaced blocks and just concatenating them. 632 SmallVector<MachineBasicBlock *, 16> Blocks; 633 Blocks.reserve(F.size()); 634 635 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 636 637 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 638 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) { 639 MachineBasicBlock *BB = &*FI; 640 Blocks.push_back(BB); 641 BlockChain &Chain = *BlockToChain[BB]; 642 if (!UpdatedPreds.insert(&Chain)) 643 continue; 644 645 assert(Chain.LoopPredecessors == 0); 646 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end(); 647 BCI != BCE; ++BCI) { 648 assert(BlockToChain[*BCI] == &Chain); 649 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(), 650 PE = (*BCI)->pred_end(); 651 PI != PE; ++PI) { 652 if (BlockToChain[*PI] == &Chain) 653 continue; 654 ++Chain.LoopPredecessors; 655 } 656 } 657 658 if (Chain.LoopPredecessors == 0) 659 BlockWorkList.push_back(BB); 660 } 661 662 BlockChain &FunctionChain = *BlockToChain[&F.front()]; 663 buildChain(&F.front(), FunctionChain, Blocks, BlockWorkList); 664 665 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType; 666 DEBUG({ 667 // Crash at the end so we get all of the debugging output first. 668 bool BadFunc = false; 669 FunctionBlockSetType FunctionBlockSet; 670 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 671 FunctionBlockSet.insert(FI); 672 673 for (BlockChain::iterator BCI = FunctionChain.begin(), 674 BCE = FunctionChain.end(); 675 BCI != BCE; ++BCI) 676 if (!FunctionBlockSet.erase(*BCI)) { 677 BadFunc = true; 678 dbgs() << "Function chain contains a block not in the function!\n" 679 << " Bad block: " << getBlockName(*BCI) << "\n"; 680 } 681 682 if (!FunctionBlockSet.empty()) { 683 BadFunc = true; 684 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(), 685 FBE = FunctionBlockSet.end(); 686 FBI != FBE; ++FBI) 687 dbgs() << "Function contains blocks never placed into a chain!\n" 688 << " Bad block: " << getBlockName(*FBI) << "\n"; 689 } 690 assert(!BadFunc && "Detected problems with the block placement."); 691 }); 692 693 // Splice the blocks into place. 694 MachineFunction::iterator InsertPos = F.begin(); 695 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch. 696 for (BlockChain::iterator BI = FunctionChain.begin(), 697 BE = FunctionChain.end(); 698 BI != BE; ++BI) { 699 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain " 700 : " ... ") 701 << getBlockName(*BI) << "\n"); 702 if (InsertPos != MachineFunction::iterator(*BI)) 703 F.splice(InsertPos, *BI); 704 else 705 ++InsertPos; 706 707 // Update the terminator of the previous block. 708 if (BI == FunctionChain.begin()) 709 continue; 710 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI)); 711 712 // FIXME: It would be awesome of updateTerminator would just return rather 713 // than assert when the branch cannot be analyzed in order to remove this 714 // boiler plate. 715 Cond.clear(); 716 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 717 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) 718 PrevBB->updateTerminator(); 719 } 720 721 // Fixup the last block. 722 Cond.clear(); 723 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 724 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond)) 725 F.back().updateTerminator(); 726} 727 728/// \brief Recursive helper to align a loop and any nested loops. 729static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) { 730 // Recurse through nested loops. 731 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I) 732 AlignLoop(F, *I, Align); 733 734 L->getTopBlock()->setAlignment(Align); 735} 736 737/// \brief Align loop headers to target preferred alignments. 738void MachineBlockPlacement::AlignLoops(MachineFunction &F) { 739 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize)) 740 return; 741 742 unsigned Align = TLI->getPrefLoopAlignment(); 743 if (!Align) 744 return; // Don't care about loop alignment. 745 746 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I) 747 AlignLoop(F, *I, Align); 748} 749 750bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) { 751 // Check for single-block functions and skip them. 752 if (llvm::next(F.begin()) == F.end()) 753 return false; 754 755 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 756 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 757 MLI = &getAnalysis<MachineLoopInfo>(); 758 TII = F.getTarget().getInstrInfo(); 759 TLI = F.getTarget().getTargetLowering(); 760 assert(BlockToChain.empty()); 761 762 buildCFGChains(F); 763 AlignLoops(F); 764 765 BlockToChain.clear(); 766 767 // We always return true as we have no way to track whether the final order 768 // differs from the original order. 769 return true; 770} 771 772namespace { 773/// \brief A pass to compute block placement statistics. 774/// 775/// A separate pass to compute interesting statistics for evaluating block 776/// placement. This is separate from the actual placement pass so that they can 777/// be computed in the absense of any placement transformations or when using 778/// alternative placement strategies. 779class MachineBlockPlacementStats : public MachineFunctionPass { 780 /// \brief A handle to the branch probability pass. 781 const MachineBranchProbabilityInfo *MBPI; 782 783 /// \brief A handle to the function-wide block frequency pass. 784 const MachineBlockFrequencyInfo *MBFI; 785 786public: 787 static char ID; // Pass identification, replacement for typeid 788 MachineBlockPlacementStats() : MachineFunctionPass(ID) { 789 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry()); 790 } 791 792 bool runOnMachineFunction(MachineFunction &F); 793 794 void getAnalysisUsage(AnalysisUsage &AU) const { 795 AU.addRequired<MachineBranchProbabilityInfo>(); 796 AU.addRequired<MachineBlockFrequencyInfo>(); 797 AU.setPreservesAll(); 798 MachineFunctionPass::getAnalysisUsage(AU); 799 } 800 801 const char *getPassName() const { return "Block Placement Stats"; } 802}; 803} 804 805char MachineBlockPlacementStats::ID = 0; 806INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats", 807 "Basic Block Placement Stats", false, false) 808INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 809INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 810INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats", 811 "Basic Block Placement Stats", false, false) 812 813FunctionPass *llvm::createMachineBlockPlacementStatsPass() { 814 return new MachineBlockPlacementStats(); 815} 816 817bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) { 818 // Check for single-block functions and skip them. 819 if (llvm::next(F.begin()) == F.end()) 820 return false; 821 822 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 823 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 824 825 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) { 826 BlockFrequency BlockFreq = MBFI->getBlockFreq(I); 827 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches 828 : NumUncondBranches; 829 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq 830 : UncondBranchTakenFreq; 831 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(), 832 SE = I->succ_end(); 833 SI != SE; ++SI) { 834 // Skip if this successor is a fallthrough. 835 if (I->isLayoutSuccessor(*SI)) 836 continue; 837 838 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI); 839 ++NumBranches; 840 BranchTakenFreq += EdgeFreq.getFrequency(); 841 } 842 } 843 844 return false; 845} 846 847