MachineBlockPlacement.cpp revision b5856c83ff4fd796c3eabccca2ed3b06173aeb51
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 BranchProbability BestProb = BranchProbability::getZero(); 338 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n"); 339 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 340 SE = BB->succ_end(); 341 SI != SE; ++SI) { 342 if (BlockFilter && !BlockFilter->count(*SI)) 343 continue; 344 BlockChain &SuccChain = *BlockToChain[*SI]; 345 if (&SuccChain == &Chain) { 346 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n"); 347 continue; 348 } 349 350 BranchProbability SuccProb = MBPI->getEdgeProbability(BB, *SI); 351 352 // Only consider successors which are either "hot", or wouldn't violate 353 // any CFG constraints. 354 if (SuccChain.LoopPredecessors != 0 && SuccProb < HotProb) { 355 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n"); 356 continue; 357 } 358 359 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb 360 << " (prob)" 361 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "") 362 << "\n"); 363 if (BestSucc && BestProb >= SuccProb) 364 continue; 365 BestSucc = *SI; 366 BestProb = SuccProb; 367 } 368 return BestSucc; 369} 370 371/// \brief Select the best block from a worklist. 372/// 373/// This looks through the provided worklist as a list of candidate basic 374/// blocks and select the most profitable one to place. The definition of 375/// profitable only really makes sense in the context of a loop. This returns 376/// the most frequently visited block in the worklist, which in the case of 377/// a loop, is the one most desirable to be physically close to the rest of the 378/// loop body in order to improve icache behavior. 379/// 380/// \returns The best block found, or null if none are viable. 381MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock( 382 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList, 383 const BlockFilterSet *BlockFilter) { 384 MachineBasicBlock *BestBlock = 0; 385 BlockFrequency BestFreq; 386 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(), 387 WBE = WorkList.end(); 388 WBI != WBE; ++WBI) { 389 if (BlockFilter && !BlockFilter->count(*WBI)) 390 continue; 391 BlockChain &SuccChain = *BlockToChain[*WBI]; 392 if (&SuccChain == &Chain) { 393 DEBUG(dbgs() << " " << getBlockName(*WBI) 394 << " -> Already merged!\n"); 395 continue; 396 } 397 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block"); 398 399 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI); 400 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq 401 << " (freq)\n"); 402 if (BestBlock && BestFreq >= CandidateFreq) 403 continue; 404 BestBlock = *WBI; 405 BestFreq = CandidateFreq; 406 } 407 return BestBlock; 408} 409 410/// \brief Retrieve the first unplaced basic block. 411/// 412/// This routine is called when we are unable to use the CFG to walk through 413/// all of the basic blocks and form a chain due to unnatural loops in the CFG. 414/// We walk through the sequence of blocks, starting from the 415/// LastUnplacedBlockIdx. We update this index to avoid re-scanning the entire 416/// sequence on repeated calls to this routine. 417MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock( 418 const BlockChain &PlacedChain, 419 ArrayRef<MachineBasicBlock *> Blocks, 420 unsigned &PrevUnplacedBlockIdx) { 421 for (unsigned i = PrevUnplacedBlockIdx, e = Blocks.size(); i != e; ++i) { 422 MachineBasicBlock *BB = Blocks[i]; 423 if (BlockToChain[BB] != &PlacedChain) { 424 PrevUnplacedBlockIdx = i; 425 return BB; 426 } 427 } 428 return 0; 429} 430 431void MachineBlockPlacement::buildChain( 432 MachineBasicBlock *BB, 433 BlockChain &Chain, 434 ArrayRef<MachineBasicBlock *> Blocks, 435 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 436 const BlockFilterSet *BlockFilter) { 437 assert(BB); 438 assert(BlockToChain[BB] == &Chain); 439 assert(*Chain.begin() == BB); 440 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch. 441 unsigned PrevUnplacedBlockIdx = 0; 442 443 MachineBasicBlock *LoopHeaderBB = BB; 444 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter); 445 BB = *llvm::prior(Chain.end()); 446 for (;;) { 447 assert(BB); 448 assert(BlockToChain[BB] == &Chain); 449 assert(*llvm::prior(Chain.end()) == BB); 450 MachineBasicBlock *BestSucc = 0; 451 452 // Check for unreasonable branches, and forcibly merge the existing layout 453 // successor for them. We can handle cases that AnalyzeBranch can't: jump 454 // tables etc are fine. The case we want to handle specially is when there 455 // is potential fallthrough, but the branch cannot be analyzed. This 456 // includes blocks without terminators as well as other cases. 457 Cond.clear(); 458 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 459 if (TII->AnalyzeBranch(*BB, TBB, FBB, Cond) && BB->canFallThrough()) { 460 MachineFunction::iterator I(BB), NextI(llvm::next(I)); 461 // Ensure that the layout successor is a viable block, as we know that 462 // fallthrough is a possibility. 463 assert(NextI != BB->getParent()->end()); 464 assert(!BlockFilter || BlockFilter->count(NextI)); 465 BestSucc = NextI; 466 } 467 468 // Otherwise, look for the best viable successor if there is one to place 469 // immediately after this block. 470 if (!BestSucc) 471 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter); 472 473 // If an immediate successor isn't available, look for the best viable 474 // block among those we've identified as not violating the loop's CFG at 475 // this point. This won't be a fallthrough, but it will increase locality. 476 if (!BestSucc) 477 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter); 478 479 if (!BestSucc) { 480 BestSucc = getFirstUnplacedBlock(Chain, Blocks, PrevUnplacedBlockIdx); 481 if (!BestSucc) 482 break; 483 484 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the " 485 "layout successor until the CFG reduces\n"); 486 } 487 488 // Place this block, updating the datastructures to reflect its placement. 489 BlockChain &SuccChain = *BlockToChain[BestSucc]; 490 // Zero out LoopPredecessors for the successor we're about to merge in case 491 // we selected a successor that didn't fit naturally into the CFG. 492 SuccChain.LoopPredecessors = 0; 493 DEBUG(dbgs() << "Merging from " << getBlockNum(BB) 494 << " to " << getBlockNum(BestSucc) << "\n"); 495 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter); 496 Chain.merge(BestSucc, &SuccChain); 497 BB = *llvm::prior(Chain.end()); 498 }; 499 500 DEBUG(dbgs() << "Finished forming chain for header block " 501 << getBlockNum(*Chain.begin()) << "\n"); 502} 503 504/// \brief Forms basic block chains from the natural loop structures. 505/// 506/// These chains are designed to preserve the existing *structure* of the code 507/// as much as possible. We can then stitch the chains together in a way which 508/// both preserves the topological structure and minimizes taken conditional 509/// branches. 510void MachineBlockPlacement::buildLoopChains(MachineFunction &F, 511 MachineLoop &L) { 512 // First recurse through any nested loops, building chains for those inner 513 // loops. 514 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI) 515 buildLoopChains(F, **LI); 516 517 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 518 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end()); 519 BlockChain &LoopChain = *BlockToChain[L.getHeader()]; 520 521 // FIXME: This is a really lame way of walking the chains in the loop: we 522 // walk the blocks, and use a set to prevent visiting a particular chain 523 // twice. 524 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 525 for (MachineLoop::block_iterator BI = L.block_begin(), 526 BE = L.block_end(); 527 BI != BE; ++BI) { 528 BlockChain &Chain = *BlockToChain[*BI]; 529 if (!UpdatedPreds.insert(&Chain) || BI == L.block_begin()) 530 continue; 531 532 assert(Chain.LoopPredecessors == 0); 533 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end(); 534 BCI != BCE; ++BCI) { 535 assert(BlockToChain[*BCI] == &Chain); 536 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(), 537 PE = (*BCI)->pred_end(); 538 PI != PE; ++PI) { 539 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI)) 540 continue; 541 ++Chain.LoopPredecessors; 542 } 543 } 544 545 if (Chain.LoopPredecessors == 0) 546 BlockWorkList.push_back(*BI); 547 } 548 549 buildChain(*L.block_begin(), LoopChain, L.getBlocks(), BlockWorkList, 550 &LoopBlockSet); 551 552 DEBUG({ 553 // Crash at the end so we get all of the debugging output first. 554 bool BadLoop = false; 555 if (LoopChain.LoopPredecessors) { 556 BadLoop = true; 557 dbgs() << "Loop chain contains a block without its preds placed!\n" 558 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 559 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"; 560 } 561 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end(); 562 BCI != BCE; ++BCI) 563 if (!LoopBlockSet.erase(*BCI)) { 564 BadLoop = true; 565 dbgs() << "Loop chain contains a block not contained by the loop!\n" 566 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 567 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 568 << " Bad block: " << getBlockName(*BCI) << "\n"; 569 } 570 571 if (!LoopBlockSet.empty()) { 572 BadLoop = true; 573 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(), 574 LBE = LoopBlockSet.end(); 575 LBI != LBE; ++LBI) 576 dbgs() << "Loop contains blocks never placed into a chain!\n" 577 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 578 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 579 << " Bad block: " << getBlockName(*LBI) << "\n"; 580 } 581 assert(!BadLoop && "Detected problems with the placement of this loop."); 582 }); 583} 584 585void MachineBlockPlacement::buildCFGChains(MachineFunction &F) { 586 // Ensure that every BB in the function has an associated chain to simplify 587 // the assumptions of the remaining algorithm. 588 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 589 BlockToChain[&*FI] = 590 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, &*FI); 591 592 // Build any loop-based chains. 593 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE; 594 ++LI) 595 buildLoopChains(F, **LI); 596 597 // We need a vector of blocks so that buildChain can handle unnatural CFG 598 // constructs by searching for unplaced blocks and just concatenating them. 599 SmallVector<MachineBasicBlock *, 16> Blocks; 600 Blocks.reserve(F.size()); 601 602 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 603 604 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 605 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) { 606 MachineBasicBlock *BB = &*FI; 607 Blocks.push_back(BB); 608 BlockChain &Chain = *BlockToChain[BB]; 609 if (!UpdatedPreds.insert(&Chain)) 610 continue; 611 612 assert(Chain.LoopPredecessors == 0); 613 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end(); 614 BCI != BCE; ++BCI) { 615 assert(BlockToChain[*BCI] == &Chain); 616 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(), 617 PE = (*BCI)->pred_end(); 618 PI != PE; ++PI) { 619 if (BlockToChain[*PI] == &Chain) 620 continue; 621 ++Chain.LoopPredecessors; 622 } 623 } 624 625 if (Chain.LoopPredecessors == 0) 626 BlockWorkList.push_back(BB); 627 } 628 629 BlockChain &FunctionChain = *BlockToChain[&F.front()]; 630 buildChain(&F.front(), FunctionChain, Blocks, BlockWorkList); 631 632 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType; 633 DEBUG({ 634 // Crash at the end so we get all of the debugging output first. 635 bool BadFunc = false; 636 FunctionBlockSetType FunctionBlockSet; 637 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 638 FunctionBlockSet.insert(FI); 639 640 for (BlockChain::iterator BCI = FunctionChain.begin(), 641 BCE = FunctionChain.end(); 642 BCI != BCE; ++BCI) 643 if (!FunctionBlockSet.erase(*BCI)) { 644 BadFunc = true; 645 dbgs() << "Function chain contains a block not in the function!\n" 646 << " Bad block: " << getBlockName(*BCI) << "\n"; 647 } 648 649 if (!FunctionBlockSet.empty()) { 650 BadFunc = true; 651 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(), 652 FBE = FunctionBlockSet.end(); 653 FBI != FBE; ++FBI) 654 dbgs() << "Function contains blocks never placed into a chain!\n" 655 << " Bad block: " << getBlockName(*FBI) << "\n"; 656 } 657 assert(!BadFunc && "Detected problems with the block placement."); 658 }); 659 660 // Splice the blocks into place. 661 MachineFunction::iterator InsertPos = F.begin(); 662 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch. 663 for (BlockChain::iterator BI = FunctionChain.begin(), 664 BE = FunctionChain.end(); 665 BI != BE; ++BI) { 666 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain " 667 : " ... ") 668 << getBlockName(*BI) << "\n"); 669 if (InsertPos != MachineFunction::iterator(*BI)) 670 F.splice(InsertPos, *BI); 671 else 672 ++InsertPos; 673 674 // Update the terminator of the previous block. 675 if (BI == FunctionChain.begin()) 676 continue; 677 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI)); 678 679 // FIXME: It would be awesome of updateTerminator would just return rather 680 // than assert when the branch cannot be analyzed in order to remove this 681 // boiler plate. 682 Cond.clear(); 683 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 684 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) 685 PrevBB->updateTerminator(); 686 } 687 688 // Fixup the last block. 689 Cond.clear(); 690 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 691 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond)) 692 F.back().updateTerminator(); 693} 694 695/// \brief Recursive helper to align a loop and any nested loops. 696static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) { 697 // Recurse through nested loops. 698 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I) 699 AlignLoop(F, *I, Align); 700 701 L->getTopBlock()->setAlignment(Align); 702} 703 704/// \brief Align loop headers to target preferred alignments. 705void MachineBlockPlacement::AlignLoops(MachineFunction &F) { 706 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize)) 707 return; 708 709 unsigned Align = TLI->getPrefLoopAlignment(); 710 if (!Align) 711 return; // Don't care about loop alignment. 712 713 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I) 714 AlignLoop(F, *I, Align); 715} 716 717bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) { 718 // Check for single-block functions and skip them. 719 if (llvm::next(F.begin()) == F.end()) 720 return false; 721 722 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 723 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 724 MLI = &getAnalysis<MachineLoopInfo>(); 725 TII = F.getTarget().getInstrInfo(); 726 TLI = F.getTarget().getTargetLowering(); 727 assert(BlockToChain.empty()); 728 729 buildCFGChains(F); 730 AlignLoops(F); 731 732 BlockToChain.clear(); 733 734 // We always return true as we have no way to track whether the final order 735 // differs from the original order. 736 return true; 737} 738 739namespace { 740/// \brief A pass to compute block placement statistics. 741/// 742/// A separate pass to compute interesting statistics for evaluating block 743/// placement. This is separate from the actual placement pass so that they can 744/// be computed in the absense of any placement transformations or when using 745/// alternative placement strategies. 746class MachineBlockPlacementStats : public MachineFunctionPass { 747 /// \brief A handle to the branch probability pass. 748 const MachineBranchProbabilityInfo *MBPI; 749 750 /// \brief A handle to the function-wide block frequency pass. 751 const MachineBlockFrequencyInfo *MBFI; 752 753public: 754 static char ID; // Pass identification, replacement for typeid 755 MachineBlockPlacementStats() : MachineFunctionPass(ID) { 756 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry()); 757 } 758 759 bool runOnMachineFunction(MachineFunction &F); 760 761 void getAnalysisUsage(AnalysisUsage &AU) const { 762 AU.addRequired<MachineBranchProbabilityInfo>(); 763 AU.addRequired<MachineBlockFrequencyInfo>(); 764 AU.setPreservesAll(); 765 MachineFunctionPass::getAnalysisUsage(AU); 766 } 767 768 const char *getPassName() const { return "Block Placement Stats"; } 769}; 770} 771 772char MachineBlockPlacementStats::ID = 0; 773INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats", 774 "Basic Block Placement Stats", false, false) 775INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 776INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 777INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats", 778 "Basic Block Placement Stats", false, false) 779 780FunctionPass *llvm::createMachineBlockPlacementStatsPass() { 781 return new MachineBlockPlacementStats(); 782} 783 784bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) { 785 // Check for single-block functions and skip them. 786 if (llvm::next(F.begin()) == F.end()) 787 return false; 788 789 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 790 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 791 792 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) { 793 BlockFrequency BlockFreq = MBFI->getBlockFreq(I); 794 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches 795 : NumUncondBranches; 796 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq 797 : UncondBranchTakenFreq; 798 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(), 799 SE = I->succ_end(); 800 SI != SE; ++SI) { 801 // Skip if this successor is a fallthrough. 802 if (I->isLayoutSuccessor(*SI)) 803 continue; 804 805 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI); 806 ++NumBranches; 807 BranchTakenFreq += EdgeFreq.getFrequency(); 808 } 809 } 810 811 return false; 812} 813 814