MachineBlockPlacement.cpp revision 10252db69bdddb445e53892b388fbe5921114b86
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 *> &Blocks, 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 void buildChain(MachineBasicBlock *BB, BlockChain &Chain, 218 SmallVectorImpl<MachineBasicBlock *> &Blocks, 219 const BlockFilterSet *BlockFilter = 0); 220 void buildLoopChains(MachineFunction &F, MachineLoop &L); 221 void buildCFGChains(MachineFunction &F); 222 void AlignLoops(MachineFunction &F); 223 224public: 225 static char ID; // Pass identification, replacement for typeid 226 MachineBlockPlacement() : MachineFunctionPass(ID) { 227 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry()); 228 } 229 230 bool runOnMachineFunction(MachineFunction &F); 231 232 void getAnalysisUsage(AnalysisUsage &AU) const { 233 AU.addRequired<MachineBranchProbabilityInfo>(); 234 AU.addRequired<MachineBlockFrequencyInfo>(); 235 AU.addRequired<MachineLoopInfo>(); 236 MachineFunctionPass::getAnalysisUsage(AU); 237 } 238 239 const char *getPassName() const { return "Block Placement"; } 240}; 241} 242 243char MachineBlockPlacement::ID = 0; 244INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2", 245 "Branch Probability Basic Block Placement", false, false) 246INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 247INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 248INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 249INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2", 250 "Branch Probability Basic Block Placement", false, false) 251 252FunctionPass *llvm::createMachineBlockPlacementPass() { 253 return new MachineBlockPlacement(); 254} 255 256#ifndef NDEBUG 257/// \brief Helper to print the name of a MBB. 258/// 259/// Only used by debug logging. 260static std::string getBlockName(MachineBasicBlock *BB) { 261 std::string Result; 262 raw_string_ostream OS(Result); 263 OS << "BB#" << BB->getNumber() 264 << " (derived from LLVM BB '" << BB->getName() << "')"; 265 OS.flush(); 266 return Result; 267} 268 269/// \brief Helper to print the number of a MBB. 270/// 271/// Only used by debug logging. 272static std::string getBlockNum(MachineBasicBlock *BB) { 273 std::string Result; 274 raw_string_ostream OS(Result); 275 OS << "BB#" << BB->getNumber(); 276 OS.flush(); 277 return Result; 278} 279#endif 280 281/// \brief Mark a chain's successors as having one fewer preds. 282/// 283/// When a chain is being merged into the "placed" chain, this routine will 284/// quickly walk the successors of each block in the chain and mark them as 285/// having one fewer active predecessor. It also adds any successors of this 286/// chain which reach the zero-predecessor state to the worklist passed in. 287void MachineBlockPlacement::markChainSuccessors( 288 BlockChain &Chain, 289 MachineBasicBlock *LoopHeaderBB, 290 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 291 const BlockFilterSet *BlockFilter) { 292 // Walk all the blocks in this chain, marking their successors as having 293 // a predecessor placed. 294 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end(); 295 CBI != CBE; ++CBI) { 296 // Add any successors for which this is the only un-placed in-loop 297 // predecessor to the worklist as a viable candidate for CFG-neutral 298 // placement. No subsequent placement of this block will violate the CFG 299 // shape, so we get to use heuristics to choose a favorable placement. 300 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(), 301 SE = (*CBI)->succ_end(); 302 SI != SE; ++SI) { 303 if (BlockFilter && !BlockFilter->count(*SI)) 304 continue; 305 BlockChain &SuccChain = *BlockToChain[*SI]; 306 // Disregard edges within a fixed chain, or edges to the loop header. 307 if (&Chain == &SuccChain || *SI == LoopHeaderBB) 308 continue; 309 310 // This is a cross-chain edge that is within the loop, so decrement the 311 // loop predecessor count of the destination chain. 312 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0) 313 BlockWorkList.push_back(*SI); 314 } 315 } 316} 317 318/// \brief Select the best successor for a block. 319/// 320/// This looks across all successors of a particular block and attempts to 321/// select the "best" one to be the layout successor. It only considers direct 322/// successors which also pass the block filter. It will attempt to avoid 323/// breaking CFG structure, but cave and break such structures in the case of 324/// very hot successor edges. 325/// 326/// \returns The best successor block found, or null if none are viable. 327MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor( 328 MachineBasicBlock *BB, BlockChain &Chain, 329 const BlockFilterSet *BlockFilter) { 330 const BranchProbability HotProb(4, 5); // 80% 331 332 MachineBasicBlock *BestSucc = 0; 333 BranchProbability BestProb = BranchProbability::getZero(); 334 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n"); 335 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 336 SE = BB->succ_end(); 337 SI != SE; ++SI) { 338 if (BlockFilter && !BlockFilter->count(*SI)) 339 continue; 340 BlockChain &SuccChain = *BlockToChain[*SI]; 341 if (&SuccChain == &Chain) { 342 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n"); 343 continue; 344 } 345 346 BranchProbability SuccProb = MBPI->getEdgeProbability(BB, *SI); 347 348 // Only consider successors which are either "hot", or wouldn't violate 349 // any CFG constraints. 350 if (SuccChain.LoopPredecessors != 0 && SuccProb < HotProb) { 351 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n"); 352 continue; 353 } 354 355 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb 356 << " (prob)" 357 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "") 358 << "\n"); 359 if (BestSucc && BestProb >= SuccProb) 360 continue; 361 BestSucc = *SI; 362 BestProb = SuccProb; 363 } 364 return BestSucc; 365} 366 367/// \brief Select the best block from a worklist. 368/// 369/// This looks through the provided worklist as a list of candidate basic 370/// blocks and select the most profitable one to place. The definition of 371/// profitable only really makes sense in the context of a loop. This returns 372/// the most frequently visited block in the worklist, which in the case of 373/// a loop, is the one most desirable to be physically close to the rest of the 374/// loop body in order to improve icache behavior. 375/// 376/// \returns The best block found, or null if none are viable. 377MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock( 378 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList, 379 const BlockFilterSet *BlockFilter) { 380 MachineBasicBlock *BestBlock = 0; 381 BlockFrequency BestFreq; 382 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(), 383 WBE = WorkList.end(); 384 WBI != WBE; ++WBI) { 385 if (BlockFilter && !BlockFilter->count(*WBI)) 386 continue; 387 BlockChain &SuccChain = *BlockToChain[*WBI]; 388 if (&SuccChain == &Chain) { 389 DEBUG(dbgs() << " " << getBlockName(*WBI) 390 << " -> Already merged!\n"); 391 continue; 392 } 393 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block"); 394 395 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI); 396 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq 397 << " (freq)\n"); 398 if (BestBlock && BestFreq >= CandidateFreq) 399 continue; 400 BestBlock = *WBI; 401 BestFreq = CandidateFreq; 402 } 403 return BestBlock; 404} 405 406void MachineBlockPlacement::buildChain( 407 MachineBasicBlock *BB, 408 BlockChain &Chain, 409 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList, 410 const BlockFilterSet *BlockFilter) { 411 assert(BB); 412 assert(BlockToChain[BB] == &Chain); 413 assert(*Chain.begin() == BB); 414 MachineBasicBlock *LoopHeaderBB = BB; 415 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter); 416 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch. 417 BB = *llvm::prior(Chain.end()); 418 for (;;) { 419 assert(BB); 420 assert(BlockToChain[BB] == &Chain); 421 assert(*llvm::prior(Chain.end()) == BB); 422 MachineBasicBlock *BestSucc = 0; 423 424 // Check for unreasonable branches, and forcibly merge the existing layout 425 // successor for them. We can handle cases that AnalyzeBranch can't: jump 426 // tables etc are fine. The case we want to handle specially is when there 427 // is potential fallthrough, but the branch cannot be analyzed. This 428 // includes blocks without terminators as well as other cases. 429 Cond.clear(); 430 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 431 if (TII->AnalyzeBranch(*BB, TBB, FBB, Cond) && BB->canFallThrough()) { 432 MachineFunction::iterator I(BB); 433 assert(llvm::next(I) != BB->getParent()->end() && 434 "The final block in the function can fallthrough!"); 435 BestSucc = llvm::next(I); 436 } 437 438 // Otherwise, look for the best viable successor if there is one to place 439 // immediately after this block. 440 if (!BestSucc) 441 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter); 442 443 if (BestSucc) { 444 // Zero out LoopPredecessors for the successor we're about to merge. We 445 // do this here instead of during the merge to catch cases where we 446 // didn't *intend* to merge despite non-zero loop predecessors. 447 BlockToChain[BestSucc]->LoopPredecessors = 0; 448 } 449 450 // If an immediate successor isn't available, look for the best viable 451 // block among those we've identified as not violating the loop's CFG at 452 // this point. This won't be a fallthrough, but it will increase locality. 453 if (!BestSucc) 454 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter); 455 456 if (!BestSucc) { 457 DEBUG(dbgs() << "Finished forming chain for header block " 458 << getBlockNum(*Chain.begin()) << "\n"); 459 return; 460 } 461 462 // Place this block, updating the datastructures to reflect its placement. 463 BlockChain &SuccChain = *BlockToChain[BestSucc]; 464 DEBUG(dbgs() << "Merging from " << getBlockNum(BB) 465 << " to " << getBlockNum(BestSucc) << "\n"); 466 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter); 467 Chain.merge(BestSucc, &SuccChain); 468 BB = *llvm::prior(Chain.end()); 469 } 470} 471 472/// \brief Forms basic block chains from the natural loop structures. 473/// 474/// These chains are designed to preserve the existing *structure* of the code 475/// as much as possible. We can then stitch the chains together in a way which 476/// both preserves the topological structure and minimizes taken conditional 477/// branches. 478void MachineBlockPlacement::buildLoopChains(MachineFunction &F, 479 MachineLoop &L) { 480 // First recurse through any nested loops, building chains for those inner 481 // loops. 482 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI) 483 buildLoopChains(F, **LI); 484 485 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 486 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end()); 487 488 // FIXME: This is a really lame way of walking the chains in the loop: we 489 // walk the blocks, and use a set to prevent visiting a particular chain 490 // twice. 491 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 492 for (MachineLoop::block_iterator BI = L.block_begin(), 493 BE = L.block_end(); 494 BI != BE; ++BI) { 495 BlockChain &Chain = *BlockToChain[*BI]; 496 if (!UpdatedPreds.insert(&Chain) || BI == L.block_begin()) 497 continue; 498 499 assert(Chain.LoopPredecessors == 0); 500 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end(); 501 BCI != BCE; ++BCI) { 502 assert(BlockToChain[*BCI] == &Chain); 503 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(), 504 PE = (*BCI)->pred_end(); 505 PI != PE; ++PI) { 506 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI)) 507 continue; 508 ++Chain.LoopPredecessors; 509 } 510 } 511 512 if (Chain.LoopPredecessors == 0) 513 BlockWorkList.push_back(*BI); 514 } 515 516 BlockChain &LoopChain = *BlockToChain[L.getHeader()]; 517 buildChain(*L.block_begin(), LoopChain, BlockWorkList, &LoopBlockSet); 518 519 DEBUG({ 520 // Crash at the end so we get all of the debugging output first. 521 bool BadLoop = false; 522 if (LoopChain.LoopPredecessors) { 523 BadLoop = true; 524 dbgs() << "Loop chain contains a block without its preds placed!\n" 525 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 526 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"; 527 } 528 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end(); 529 BCI != BCE; ++BCI) 530 if (!LoopBlockSet.erase(*BCI)) { 531 BadLoop = true; 532 dbgs() << "Loop chain contains a block not contained by the loop!\n" 533 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 534 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 535 << " Bad block: " << getBlockName(*BCI) << "\n"; 536 } 537 538 if (!LoopBlockSet.empty()) { 539 BadLoop = true; 540 for (SmallPtrSet<MachineBasicBlock *, 16>::iterator LBI = LoopBlockSet.begin(), LBE = LoopBlockSet.end(); 541 LBI != LBE; ++LBI) 542 dbgs() << "Loop contains blocks never placed into a chain!\n" 543 << " Loop header: " << getBlockName(*L.block_begin()) << "\n" 544 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n" 545 << " Bad block: " << getBlockName(*LBI) << "\n"; 546 } 547 assert(!BadLoop && "Detected problems with the placement of this loop."); 548 }); 549} 550 551void MachineBlockPlacement::buildCFGChains(MachineFunction &F) { 552 // Ensure that every BB in the function has an associated chain to simplify 553 // the assumptions of the remaining algorithm. 554 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 555 BlockToChain[&*FI] = 556 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, &*FI); 557 558 // Build any loop-based chains. 559 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE; 560 ++LI) 561 buildLoopChains(F, **LI); 562 563 SmallVector<MachineBasicBlock *, 16> BlockWorkList; 564 565 SmallPtrSet<BlockChain *, 4> UpdatedPreds; 566 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) { 567 MachineBasicBlock *BB = &*FI; 568 BlockChain &Chain = *BlockToChain[BB]; 569 if (!UpdatedPreds.insert(&Chain)) 570 continue; 571 572 assert(Chain.LoopPredecessors == 0); 573 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end(); 574 BCI != BCE; ++BCI) { 575 assert(BlockToChain[*BCI] == &Chain); 576 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(), 577 PE = (*BCI)->pred_end(); 578 PI != PE; ++PI) { 579 if (BlockToChain[*PI] == &Chain) 580 continue; 581 ++Chain.LoopPredecessors; 582 } 583 } 584 585 if (Chain.LoopPredecessors == 0) 586 BlockWorkList.push_back(BB); 587 } 588 589 BlockChain &FunctionChain = *BlockToChain[&F.front()]; 590 buildChain(&F.front(), FunctionChain, BlockWorkList); 591 592 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType; 593 DEBUG({ 594 // Crash at the end so we get all of the debugging output first. 595 bool BadFunc = false; 596 FunctionBlockSetType FunctionBlockSet; 597 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 598 FunctionBlockSet.insert(FI); 599 600 for (BlockChain::iterator BCI = FunctionChain.begin(), BCE = FunctionChain.end(); 601 BCI != BCE; ++BCI) 602 if (!FunctionBlockSet.erase(*BCI)) { 603 BadFunc = true; 604 dbgs() << "Function chain contains a block not in the function!\n" 605 << " Bad block: " << getBlockName(*BCI) << "\n"; 606 } 607 608 if (!FunctionBlockSet.empty()) { 609 BadFunc = true; 610 for (SmallPtrSet<MachineBasicBlock *, 16>::iterator FBI = FunctionBlockSet.begin(), 611 FBE = FunctionBlockSet.end(); FBI != FBE; ++FBI) 612 dbgs() << "Function contains blocks never placed into a chain!\n" 613 << " Bad block: " << getBlockName(*FBI) << "\n"; 614 } 615 assert(!BadFunc && "Detected problems with the block placement."); 616 }); 617 618 // Splice the blocks into place. 619 MachineFunction::iterator InsertPos = F.begin(); 620 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch. 621 for (BlockChain::iterator BI = FunctionChain.begin(), BE = FunctionChain.end(); 622 BI != BE; ++BI) { 623 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain " 624 : " ... ") 625 << getBlockName(*BI) << "\n"); 626 if (InsertPos != MachineFunction::iterator(*BI)) 627 F.splice(InsertPos, *BI); 628 else 629 ++InsertPos; 630 631 // Update the terminator of the previous block. 632 if (BI == FunctionChain.begin()) 633 continue; 634 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI)); 635 636 // FIXME: It would be awesome of updateTerminator would just return rather 637 // than assert when the branch cannot be analyzed in order to remove this 638 // boiler plate. 639 Cond.clear(); 640 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 641 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) 642 PrevBB->updateTerminator(); 643 } 644 645 // Fixup the last block. 646 Cond.clear(); 647 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch. 648 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond)) 649 F.back().updateTerminator(); 650} 651 652/// \brief Recursive helper to align a loop and any nested loops. 653static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) { 654 // Recurse through nested loops. 655 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I) 656 AlignLoop(F, *I, Align); 657 658 L->getTopBlock()->setAlignment(Align); 659} 660 661/// \brief Align loop headers to target preferred alignments. 662void MachineBlockPlacement::AlignLoops(MachineFunction &F) { 663 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize)) 664 return; 665 666 unsigned Align = TLI->getPrefLoopAlignment(); 667 if (!Align) 668 return; // Don't care about loop alignment. 669 670 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I) 671 AlignLoop(F, *I, Align); 672} 673 674bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) { 675 // Check for single-block functions and skip them. 676 if (llvm::next(F.begin()) == F.end()) 677 return false; 678 679 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 680 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 681 MLI = &getAnalysis<MachineLoopInfo>(); 682 TII = F.getTarget().getInstrInfo(); 683 TLI = F.getTarget().getTargetLowering(); 684 assert(BlockToChain.empty()); 685 686 buildCFGChains(F); 687 AlignLoops(F); 688 689 BlockToChain.clear(); 690 691 // We always return true as we have no way to track whether the final order 692 // differs from the original order. 693 return true; 694} 695 696namespace { 697/// \brief A pass to compute block placement statistics. 698/// 699/// A separate pass to compute interesting statistics for evaluating block 700/// placement. This is separate from the actual placement pass so that they can 701/// be computed in the absense of any placement transformations or when using 702/// alternative placement strategies. 703class MachineBlockPlacementStats : public MachineFunctionPass { 704 /// \brief A handle to the branch probability pass. 705 const MachineBranchProbabilityInfo *MBPI; 706 707 /// \brief A handle to the function-wide block frequency pass. 708 const MachineBlockFrequencyInfo *MBFI; 709 710public: 711 static char ID; // Pass identification, replacement for typeid 712 MachineBlockPlacementStats() : MachineFunctionPass(ID) { 713 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry()); 714 } 715 716 bool runOnMachineFunction(MachineFunction &F); 717 718 void getAnalysisUsage(AnalysisUsage &AU) const { 719 AU.addRequired<MachineBranchProbabilityInfo>(); 720 AU.addRequired<MachineBlockFrequencyInfo>(); 721 AU.setPreservesAll(); 722 MachineFunctionPass::getAnalysisUsage(AU); 723 } 724 725 const char *getPassName() const { return "Block Placement Stats"; } 726}; 727} 728 729char MachineBlockPlacementStats::ID = 0; 730INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats", 731 "Basic Block Placement Stats", false, false) 732INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 733INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 734INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats", 735 "Basic Block Placement Stats", false, false) 736 737FunctionPass *llvm::createMachineBlockPlacementStatsPass() { 738 return new MachineBlockPlacementStats(); 739} 740 741bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) { 742 // Check for single-block functions and skip them. 743 if (llvm::next(F.begin()) == F.end()) 744 return false; 745 746 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 747 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 748 749 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) { 750 BlockFrequency BlockFreq = MBFI->getBlockFreq(I); 751 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches 752 : NumUncondBranches; 753 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq 754 : UncondBranchTakenFreq; 755 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(), 756 SE = I->succ_end(); 757 SI != SE; ++SI) { 758 // Skip if this successor is a fallthrough. 759 if (I->isLayoutSuccessor(*SI)) 760 continue; 761 762 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI); 763 ++NumBranches; 764 BranchTakenFreq += EdgeFreq.getFrequency(); 765 } 766 } 767 768 return false; 769} 770 771