LoopUnswitch.cpp revision 831737d329a727f53a1fb0572f7b7a8127208881
1//===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===// 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 pass transforms loops that contain branches on loop-invariant conditions 11// to have multiple loops. For example, it turns the left into the right code: 12// 13// for (...) if (lic) 14// A for (...) 15// if (lic) A; B; C 16// B else 17// C for (...) 18// A; C 19// 20// This can increase the size of the code exponentially (doubling it every time 21// a loop is unswitched) so we only unswitch if the resultant code will be 22// smaller than a threshold. 23// 24// This pass expects LICM to be run before it to hoist invariant conditions out 25// of the loop, to make the unswitching opportunity obvious. 26// 27//===----------------------------------------------------------------------===// 28 29#define DEBUG_TYPE "loop-unswitch" 30#include "llvm/Transforms/Scalar.h" 31#include "llvm/ADT/STLExtras.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/Analysis/CodeMetrics.h" 35#include "llvm/Analysis/Dominators.h" 36#include "llvm/Analysis/InstructionSimplify.h" 37#include "llvm/Analysis/LoopInfo.h" 38#include "llvm/Analysis/LoopPass.h" 39#include "llvm/Analysis/ScalarEvolution.h" 40#include "llvm/Constants.h" 41#include "llvm/DerivedTypes.h" 42#include "llvm/Function.h" 43#include "llvm/Instructions.h" 44#include "llvm/Support/CommandLine.h" 45#include "llvm/Support/Debug.h" 46#include "llvm/Support/raw_ostream.h" 47#include "llvm/Transforms/Utils/BasicBlockUtils.h" 48#include "llvm/Transforms/Utils/Cloning.h" 49#include "llvm/Transforms/Utils/Local.h" 50#include <algorithm> 51#include <map> 52#include <set> 53using namespace llvm; 54 55STATISTIC(NumBranches, "Number of branches unswitched"); 56STATISTIC(NumSwitches, "Number of switches unswitched"); 57STATISTIC(NumSelects , "Number of selects unswitched"); 58STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 59STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 60STATISTIC(TotalInsts, "Total number of instructions analyzed"); 61 62// The specific value of 100 here was chosen based only on intuition and a 63// few specific examples. 64static cl::opt<unsigned> 65Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 66 cl::init(100), cl::Hidden); 67 68namespace { 69 70 class LUAnalysisCache { 71 72 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> > 73 UnswitchedValsMap; 74 75 typedef UnswitchedValsMap::iterator UnswitchedValsIt; 76 77 struct LoopProperties { 78 unsigned CanBeUnswitchedCount; 79 unsigned SizeEstimation; 80 UnswitchedValsMap UnswitchedVals; 81 }; 82 83 // Here we use std::map instead of DenseMap, since we need to keep valid 84 // LoopProperties pointer for current loop for better performance. 85 typedef std::map<const Loop*, LoopProperties> LoopPropsMap; 86 typedef LoopPropsMap::iterator LoopPropsMapIt; 87 88 LoopPropsMap LoopsProperties; 89 UnswitchedValsMap* CurLoopInstructions; 90 LoopProperties* CurrentLoopProperties; 91 92 // Max size of code we can produce on remained iterations. 93 unsigned MaxSize; 94 95 public: 96 97 LUAnalysisCache() : 98 CurLoopInstructions(NULL), CurrentLoopProperties(NULL), 99 MaxSize(Threshold) 100 {} 101 102 // Analyze loop. Check its size, calculate is it possible to unswitch 103 // it. Returns true if we can unswitch this loop. 104 bool countLoop(const Loop* L); 105 106 // Clean all data related to given loop. 107 void forgetLoop(const Loop* L); 108 109 // Mark case value as unswitched. 110 // Since SI instruction can be partly unswitched, in order to avoid 111 // extra unswitching in cloned loops keep track all unswitched values. 112 void setUnswitched(const SwitchInst* SI, const Value* V); 113 114 // Check was this case value unswitched before or not. 115 bool isUnswitched(const SwitchInst* SI, const Value* V); 116 117 // Clone all loop-unswitch related loop properties. 118 // Redistribute unswitching quotas. 119 // Note, that new loop data is stored inside the VMap. 120 void cloneData(const Loop* NewLoop, const Loop* OldLoop, 121 const ValueToValueMapTy& VMap); 122 }; 123 124 class LoopUnswitch : public LoopPass { 125 LoopInfo *LI; // Loop information 126 LPPassManager *LPM; 127 128 // LoopProcessWorklist - Used to check if second loop needs processing 129 // after RewriteLoopBodyWithConditionConstant rewrites first loop. 130 std::vector<Loop*> LoopProcessWorklist; 131 132 LUAnalysisCache BranchesInfo; 133 134 bool OptimizeForSize; 135 bool redoLoop; 136 137 Loop *currentLoop; 138 DominatorTree *DT; 139 BasicBlock *loopHeader; 140 BasicBlock *loopPreheader; 141 142 // LoopBlocks contains all of the basic blocks of the loop, including the 143 // preheader of the loop, the body of the loop, and the exit blocks of the 144 // loop, in that order. 145 std::vector<BasicBlock*> LoopBlocks; 146 // NewBlocks contained cloned copy of basic blocks from LoopBlocks. 147 std::vector<BasicBlock*> NewBlocks; 148 149 public: 150 static char ID; // Pass ID, replacement for typeid 151 explicit LoopUnswitch(bool Os = false) : 152 LoopPass(ID), OptimizeForSize(Os), redoLoop(false), 153 currentLoop(NULL), DT(NULL), loopHeader(NULL), 154 loopPreheader(NULL) { 155 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry()); 156 } 157 158 bool runOnLoop(Loop *L, LPPassManager &LPM); 159 bool processCurrentLoop(); 160 161 /// This transformation requires natural loop information & requires that 162 /// loop preheaders be inserted into the CFG. 163 /// 164 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 165 AU.addRequiredID(LoopSimplifyID); 166 AU.addPreservedID(LoopSimplifyID); 167 AU.addRequired<LoopInfo>(); 168 AU.addPreserved<LoopInfo>(); 169 AU.addRequiredID(LCSSAID); 170 AU.addPreservedID(LCSSAID); 171 AU.addPreserved<DominatorTree>(); 172 AU.addPreserved<ScalarEvolution>(); 173 } 174 175 private: 176 177 virtual void releaseMemory() { 178 BranchesInfo.forgetLoop(currentLoop); 179 } 180 181 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 182 /// remove it. 183 void RemoveLoopFromWorklist(Loop *L) { 184 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 185 LoopProcessWorklist.end(), L); 186 if (I != LoopProcessWorklist.end()) 187 LoopProcessWorklist.erase(I); 188 } 189 190 void initLoopData() { 191 loopHeader = currentLoop->getHeader(); 192 loopPreheader = currentLoop->getLoopPreheader(); 193 } 194 195 /// Split all of the edges from inside the loop to their exit blocks. 196 /// Update the appropriate Phi nodes as we do so. 197 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks); 198 199 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); 200 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 201 BasicBlock *ExitBlock); 202 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 203 204 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 205 Constant *Val, bool isEqual); 206 207 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 208 BasicBlock *TrueDest, 209 BasicBlock *FalseDest, 210 Instruction *InsertPt); 211 212 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 213 void RemoveBlockIfDead(BasicBlock *BB, 214 std::vector<Instruction*> &Worklist, Loop *l); 215 void RemoveLoopFromHierarchy(Loop *L); 216 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, 217 BasicBlock **LoopExit = 0); 218 219 }; 220} 221 222// Analyze loop. Check its size, calculate is it possible to unswitch 223// it. Returns true if we can unswitch this loop. 224bool LUAnalysisCache::countLoop(const Loop* L) { 225 226 std::pair<LoopPropsMapIt, bool> InsertRes = 227 LoopsProperties.insert(std::make_pair(L, LoopProperties())); 228 229 LoopProperties& Props = InsertRes.first->second; 230 231 if (InsertRes.second) { 232 // New loop. 233 234 // Limit the number of instructions to avoid causing significant code 235 // expansion, and the number of basic blocks, to avoid loops with 236 // large numbers of branches which cause loop unswitching to go crazy. 237 // This is a very ad-hoc heuristic. 238 239 // FIXME: This is overly conservative because it does not take into 240 // consideration code simplification opportunities and code that can 241 // be shared by the resultant unswitched loops. 242 CodeMetrics Metrics; 243 for (Loop::block_iterator I = L->block_begin(), 244 E = L->block_end(); 245 I != E; ++I) 246 Metrics.analyzeBasicBlock(*I); 247 248 Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5); 249 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation); 250 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount; 251 252 if (Metrics.notDuplicatable) { 253 DEBUG(dbgs() << "NOT unswitching loop %" 254 << L->getHeader()->getName() << ", contents cannot be " 255 << "duplicated!\n"); 256 return false; 257 } 258 } 259 260 if (!Props.CanBeUnswitchedCount) { 261 DEBUG(dbgs() << "NOT unswitching loop %" 262 << L->getHeader()->getName() << ", cost too high: " 263 << L->getBlocks().size() << "\n"); 264 265 return false; 266 } 267 268 // Be careful. This links are good only before new loop addition. 269 CurrentLoopProperties = &Props; 270 CurLoopInstructions = &Props.UnswitchedVals; 271 272 return true; 273} 274 275// Clean all data related to given loop. 276void LUAnalysisCache::forgetLoop(const Loop* L) { 277 278 LoopPropsMapIt LIt = LoopsProperties.find(L); 279 280 if (LIt != LoopsProperties.end()) { 281 LoopProperties& Props = LIt->second; 282 MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation; 283 LoopsProperties.erase(LIt); 284 } 285 286 CurrentLoopProperties = NULL; 287 CurLoopInstructions = NULL; 288} 289 290// Mark case value as unswitched. 291// Since SI instruction can be partly unswitched, in order to avoid 292// extra unswitching in cloned loops keep track all unswitched values. 293void LUAnalysisCache::setUnswitched(const SwitchInst* SI, const Value* V) { 294 (*CurLoopInstructions)[SI].insert(V); 295} 296 297// Check was this case value unswitched before or not. 298bool LUAnalysisCache::isUnswitched(const SwitchInst* SI, const Value* V) { 299 return (*CurLoopInstructions)[SI].count(V); 300} 301 302// Clone all loop-unswitch related loop properties. 303// Redistribute unswitching quotas. 304// Note, that new loop data is stored inside the VMap. 305void LUAnalysisCache::cloneData(const Loop* NewLoop, const Loop* OldLoop, 306 const ValueToValueMapTy& VMap) { 307 308 LoopProperties& NewLoopProps = LoopsProperties[NewLoop]; 309 LoopProperties& OldLoopProps = *CurrentLoopProperties; 310 UnswitchedValsMap& Insts = OldLoopProps.UnswitchedVals; 311 312 // Reallocate "can-be-unswitched quota" 313 314 --OldLoopProps.CanBeUnswitchedCount; 315 unsigned Quota = OldLoopProps.CanBeUnswitchedCount; 316 NewLoopProps.CanBeUnswitchedCount = Quota / 2; 317 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2; 318 319 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation; 320 321 // Clone unswitched values info: 322 // for new loop switches we clone info about values that was 323 // already unswitched and has redundant successors. 324 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) { 325 const SwitchInst* OldInst = I->first; 326 Value* NewI = VMap.lookup(OldInst); 327 const SwitchInst* NewInst = cast_or_null<SwitchInst>(NewI); 328 assert(NewInst && "All instructions that are in SrcBB must be in VMap."); 329 330 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst]; 331 } 332} 333 334char LoopUnswitch::ID = 0; 335INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", 336 false, false) 337INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 338INITIALIZE_PASS_DEPENDENCY(LoopInfo) 339INITIALIZE_PASS_DEPENDENCY(LCSSA) 340INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", 341 false, false) 342 343Pass *llvm::createLoopUnswitchPass(bool Os) { 344 return new LoopUnswitch(Os); 345} 346 347/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 348/// invariant in the loop, or has an invariant piece, return the invariant. 349/// Otherwise, return null. 350static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 351 352 // We started analyze new instruction, increment scanned instructions counter. 353 ++TotalInsts; 354 355 // We can never unswitch on vector conditions. 356 if (Cond->getType()->isVectorTy()) 357 return 0; 358 359 // Constants should be folded, not unswitched on! 360 if (isa<Constant>(Cond)) return 0; 361 362 // TODO: Handle: br (VARIANT|INVARIANT). 363 364 // Hoist simple values out. 365 if (L->makeLoopInvariant(Cond, Changed)) 366 return Cond; 367 368 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 369 if (BO->getOpcode() == Instruction::And || 370 BO->getOpcode() == Instruction::Or) { 371 // If either the left or right side is invariant, we can unswitch on this, 372 // which will cause the branch to go away in one loop and the condition to 373 // simplify in the other one. 374 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 375 return LHS; 376 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 377 return RHS; 378 } 379 380 return 0; 381} 382 383bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 384 LI = &getAnalysis<LoopInfo>(); 385 LPM = &LPM_Ref; 386 DT = getAnalysisIfAvailable<DominatorTree>(); 387 currentLoop = L; 388 Function *F = currentLoop->getHeader()->getParent(); 389 bool Changed = false; 390 do { 391 assert(currentLoop->isLCSSAForm(*DT)); 392 redoLoop = false; 393 Changed |= processCurrentLoop(); 394 } while(redoLoop); 395 396 if (Changed) { 397 // FIXME: Reconstruct dom info, because it is not preserved properly. 398 if (DT) 399 DT->runOnFunction(*F); 400 } 401 return Changed; 402} 403 404/// processCurrentLoop - Do actual work and unswitch loop if possible 405/// and profitable. 406bool LoopUnswitch::processCurrentLoop() { 407 bool Changed = false; 408 409 initLoopData(); 410 411 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 412 if (!loopPreheader) 413 return false; 414 415 // Loops with indirectbr cannot be cloned. 416 if (!currentLoop->isSafeToClone()) 417 return false; 418 419 // Without dedicated exits, splitting the exit edge may fail. 420 if (!currentLoop->hasDedicatedExits()) 421 return false; 422 423 LLVMContext &Context = loopHeader->getContext(); 424 425 // Probably we reach the quota of branches for this loop. If so 426 // stop unswitching. 427 if (!BranchesInfo.countLoop(currentLoop)) 428 return false; 429 430 // Loop over all of the basic blocks in the loop. If we find an interior 431 // block that is branching on a loop-invariant condition, we can unswitch this 432 // loop. 433 for (Loop::block_iterator I = currentLoop->block_begin(), 434 E = currentLoop->block_end(); I != E; ++I) { 435 TerminatorInst *TI = (*I)->getTerminator(); 436 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 437 // If this isn't branching on an invariant condition, we can't unswitch 438 // it. 439 if (BI->isConditional()) { 440 // See if this, or some part of it, is loop invariant. If so, we can 441 // unswitch on it if we desire. 442 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 443 currentLoop, Changed); 444 if (LoopCond && UnswitchIfProfitable(LoopCond, 445 ConstantInt::getTrue(Context))) { 446 ++NumBranches; 447 return true; 448 } 449 } 450 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 451 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 452 currentLoop, Changed); 453 unsigned NumCases = SI->getNumCases(); 454 if (LoopCond && NumCases) { 455 // Find a value to unswitch on: 456 // FIXME: this should chose the most expensive case! 457 // FIXME: scan for a case with a non-critical edge? 458 Constant *UnswitchVal = NULL; 459 460 // Do not process same value again and again. 461 // At this point we have some cases already unswitched and 462 // some not yet unswitched. Let's find the first not yet unswitched one. 463 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 464 i != e; ++i) { 465 Constant* UnswitchValCandidate = i.getCaseValue(); 466 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) { 467 UnswitchVal = UnswitchValCandidate; 468 break; 469 } 470 } 471 472 if (!UnswitchVal) 473 continue; 474 475 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 476 ++NumSwitches; 477 return true; 478 } 479 } 480 } 481 482 // Scan the instructions to check for unswitchable values. 483 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 484 BBI != E; ++BBI) 485 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 486 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 487 currentLoop, Changed); 488 if (LoopCond && UnswitchIfProfitable(LoopCond, 489 ConstantInt::getTrue(Context))) { 490 ++NumSelects; 491 return true; 492 } 493 } 494 } 495 return Changed; 496} 497 498/// isTrivialLoopExitBlock - Check to see if all paths from BB exit the 499/// loop with no side effects (including infinite loops). 500/// 501/// If true, we return true and set ExitBB to the block we 502/// exit through. 503/// 504static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 505 BasicBlock *&ExitBB, 506 std::set<BasicBlock*> &Visited) { 507 if (!Visited.insert(BB).second) { 508 // Already visited. Without more analysis, this could indicate an infinite 509 // loop. 510 return false; 511 } else if (!L->contains(BB)) { 512 // Otherwise, this is a loop exit, this is fine so long as this is the 513 // first exit. 514 if (ExitBB != 0) return false; 515 ExitBB = BB; 516 return true; 517 } 518 519 // Otherwise, this is an unvisited intra-loop node. Check all successors. 520 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 521 // Check to see if the successor is a trivial loop exit. 522 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 523 return false; 524 } 525 526 // Okay, everything after this looks good, check to make sure that this block 527 // doesn't include any side effects. 528 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 529 if (I->mayHaveSideEffects()) 530 return false; 531 532 return true; 533} 534 535/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 536/// leads to an exit from the specified loop, and has no side-effects in the 537/// process. If so, return the block that is exited to, otherwise return null. 538static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 539 std::set<BasicBlock*> Visited; 540 Visited.insert(L->getHeader()); // Branches to header make infinite loops. 541 BasicBlock *ExitBB = 0; 542 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 543 return ExitBB; 544 return 0; 545} 546 547/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 548/// trivial: that is, that the condition controls whether or not the loop does 549/// anything at all. If this is a trivial condition, unswitching produces no 550/// code duplications (equivalently, it produces a simpler loop and a new empty 551/// loop, which gets deleted). 552/// 553/// If this is a trivial condition, return true, otherwise return false. When 554/// returning true, this sets Cond and Val to the condition that controls the 555/// trivial condition: when Cond dynamically equals Val, the loop is known to 556/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 557/// Cond == Val. 558/// 559bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 560 BasicBlock **LoopExit) { 561 BasicBlock *Header = currentLoop->getHeader(); 562 TerminatorInst *HeaderTerm = Header->getTerminator(); 563 LLVMContext &Context = Header->getContext(); 564 565 BasicBlock *LoopExitBB = 0; 566 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 567 // If the header block doesn't end with a conditional branch on Cond, we 568 // can't handle it. 569 if (!BI->isConditional() || BI->getCondition() != Cond) 570 return false; 571 572 // Check to see if a successor of the branch is guaranteed to 573 // exit through a unique exit block without having any 574 // side-effects. If so, determine the value of Cond that causes it to do 575 // this. 576 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 577 BI->getSuccessor(0)))) { 578 if (Val) *Val = ConstantInt::getTrue(Context); 579 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 580 BI->getSuccessor(1)))) { 581 if (Val) *Val = ConstantInt::getFalse(Context); 582 } 583 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 584 // If this isn't a switch on Cond, we can't handle it. 585 if (SI->getCondition() != Cond) return false; 586 587 // Check to see if a successor of the switch is guaranteed to go to the 588 // latch block or exit through a one exit block without having any 589 // side-effects. If so, determine the value of Cond that causes it to do 590 // this. 591 // Note that we can't trivially unswitch on the default case or 592 // on already unswitched cases. 593 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 594 i != e; ++i) { 595 BasicBlock* LoopExitCandidate; 596 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop, 597 i.getCaseSuccessor()))) { 598 // Okay, we found a trivial case, remember the value that is trivial. 599 ConstantInt* CaseVal = i.getCaseValue(); 600 601 // Check that it was not unswitched before, since already unswitched 602 // trivial vals are looks trivial too. 603 if (BranchesInfo.isUnswitched(SI, CaseVal)) 604 continue; 605 LoopExitBB = LoopExitCandidate; 606 if (Val) *Val = CaseVal; 607 break; 608 } 609 } 610 } 611 612 // If we didn't find a single unique LoopExit block, or if the loop exit block 613 // contains phi nodes, this isn't trivial. 614 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 615 return false; // Can't handle this. 616 617 if (LoopExit) *LoopExit = LoopExitBB; 618 619 // We already know that nothing uses any scalar values defined inside of this 620 // loop. As such, we just have to check to see if this loop will execute any 621 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 622 // part of the loop that the code *would* execute. We already checked the 623 // tail, check the header now. 624 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 625 if (I->mayHaveSideEffects()) 626 return false; 627 return true; 628} 629 630/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 631/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 632/// unswitch the loop, reprocess the pieces, then return true. 633bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { 634 Function *F = loopHeader->getParent(); 635 Constant *CondVal = 0; 636 BasicBlock *ExitBlock = 0; 637 638 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 639 // If the condition is trivial, always unswitch. There is no code growth 640 // for this case. 641 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 642 return true; 643 } 644 645 // Check to see if it would be profitable to unswitch current loop. 646 647 // Do not do non-trivial unswitch while optimizing for size. 648 if (OptimizeForSize || 649 F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 650 Attribute::OptimizeForSize)) 651 return false; 652 653 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 654 return true; 655} 656 657/// CloneLoop - Recursively clone the specified loop and all of its children, 658/// mapping the blocks with the specified map. 659static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, 660 LoopInfo *LI, LPPassManager *LPM) { 661 Loop *New = new Loop(); 662 LPM->insertLoop(New, PL); 663 664 // Add all of the blocks in L to the new loop. 665 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 666 I != E; ++I) 667 if (LI->getLoopFor(*I) == L) 668 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 669 670 // Add all of the subloops to the new loop. 671 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 672 CloneLoop(*I, New, VM, LI, LPM); 673 674 return New; 675} 676 677/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 678/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 679/// code immediately before InsertPt. 680void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 681 BasicBlock *TrueDest, 682 BasicBlock *FalseDest, 683 Instruction *InsertPt) { 684 // Insert a conditional branch on LIC to the two preheaders. The original 685 // code is the true version and the new code is the false version. 686 Value *BranchVal = LIC; 687 if (!isa<ConstantInt>(Val) || 688 Val->getType() != Type::getInt1Ty(LIC->getContext())) 689 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val); 690 else if (Val != ConstantInt::getTrue(Val->getContext())) 691 // We want to enter the new loop when the condition is true. 692 std::swap(TrueDest, FalseDest); 693 694 // Insert the new branch. 695 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 696 697 // If either edge is critical, split it. This helps preserve LoopSimplify 698 // form for enclosing loops. 699 SplitCriticalEdge(BI, 0, this, false, false, true); 700 SplitCriticalEdge(BI, 1, this, false, false, true); 701} 702 703/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 704/// condition in it (a cond branch from its header block to its latch block, 705/// where the path through the loop that doesn't execute its body has no 706/// side-effects), unswitch it. This doesn't involve any code duplication, just 707/// moving the conditional branch outside of the loop and updating loop info. 708void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 709 Constant *Val, 710 BasicBlock *ExitBlock) { 711 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 712 << loopHeader->getName() << " [" << L->getBlocks().size() 713 << " blocks] in Function " << L->getHeader()->getParent()->getName() 714 << " on cond: " << *Val << " == " << *Cond << "\n"); 715 716 // First step, split the preheader, so that we know that there is a safe place 717 // to insert the conditional branch. We will change loopPreheader to have a 718 // conditional branch on Cond. 719 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 720 721 // Now that we have a place to insert the conditional branch, create a place 722 // to branch to: this is the exit block out of the loop that we should 723 // short-circuit to. 724 725 // Split this block now, so that the loop maintains its exit block, and so 726 // that the jump from the preheader can execute the contents of the exit block 727 // without actually branching to it (the exit block should be dominated by the 728 // loop header, not the preheader). 729 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 730 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 731 732 // Okay, now we have a position to branch from and a position to branch to, 733 // insert the new conditional branch. 734 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 735 loopPreheader->getTerminator()); 736 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 737 loopPreheader->getTerminator()->eraseFromParent(); 738 739 // We need to reprocess this loop, it could be unswitched again. 740 redoLoop = true; 741 742 // Now that we know that the loop is never entered when this condition is a 743 // particular value, rewrite the loop with this info. We know that this will 744 // at least eliminate the old branch. 745 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 746 ++NumTrivial; 747} 748 749/// SplitExitEdges - Split all of the edges from inside the loop to their exit 750/// blocks. Update the appropriate Phi nodes as we do so. 751void LoopUnswitch::SplitExitEdges(Loop *L, 752 const SmallVector<BasicBlock *, 8> &ExitBlocks){ 753 754 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 755 BasicBlock *ExitBlock = ExitBlocks[i]; 756 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 757 pred_end(ExitBlock)); 758 759 // Although SplitBlockPredecessors doesn't preserve loop-simplify in 760 // general, if we call it on all predecessors of all exits then it does. 761 if (!ExitBlock->isLandingPad()) { 762 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this); 763 } else { 764 SmallVector<BasicBlock*, 2> NewBBs; 765 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa", 766 this, NewBBs); 767 } 768 } 769} 770 771/// UnswitchNontrivialCondition - We determined that the loop is profitable 772/// to unswitch when LIC equal Val. Split it into loop versions and test the 773/// condition outside of either loop. Return the loops created as Out1/Out2. 774void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 775 Loop *L) { 776 Function *F = loopHeader->getParent(); 777 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 778 << loopHeader->getName() << " [" << L->getBlocks().size() 779 << " blocks] in Function " << F->getName() 780 << " when '" << *Val << "' == " << *LIC << "\n"); 781 782 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>()) 783 SE->forgetLoop(L); 784 785 LoopBlocks.clear(); 786 NewBlocks.clear(); 787 788 // First step, split the preheader and exit blocks, and add these blocks to 789 // the LoopBlocks list. 790 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 791 LoopBlocks.push_back(NewPreheader); 792 793 // We want the loop to come after the preheader, but before the exit blocks. 794 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 795 796 SmallVector<BasicBlock*, 8> ExitBlocks; 797 L->getUniqueExitBlocks(ExitBlocks); 798 799 // Split all of the edges from inside the loop to their exit blocks. Update 800 // the appropriate Phi nodes as we do so. 801 SplitExitEdges(L, ExitBlocks); 802 803 // The exit blocks may have been changed due to edge splitting, recompute. 804 ExitBlocks.clear(); 805 L->getUniqueExitBlocks(ExitBlocks); 806 807 // Add exit blocks to the loop blocks. 808 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 809 810 // Next step, clone all of the basic blocks that make up the loop (including 811 // the loop preheader and exit blocks), keeping track of the mapping between 812 // the instructions and blocks. 813 NewBlocks.reserve(LoopBlocks.size()); 814 ValueToValueMapTy VMap; 815 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 816 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); 817 818 NewBlocks.push_back(NewBB); 819 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. 820 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); 821 } 822 823 // Splice the newly inserted blocks into the function right before the 824 // original preheader. 825 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(), 826 NewBlocks[0], F->end()); 827 828 // Now we create the new Loop object for the versioned loop. 829 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); 830 831 // Recalculate unswitching quota, inherit simplified switches info for NewBB, 832 // Probably clone more loop-unswitch related loop properties. 833 BranchesInfo.cloneData(NewLoop, L, VMap); 834 835 Loop *ParentLoop = L->getParentLoop(); 836 if (ParentLoop) { 837 // Make sure to add the cloned preheader and exit blocks to the parent loop 838 // as well. 839 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 840 } 841 842 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 843 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]); 844 // The new exit block should be in the same loop as the old one. 845 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 846 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 847 848 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 849 "Exit block should have been split to have one successor!"); 850 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 851 852 // If the successor of the exit block had PHI nodes, add an entry for 853 // NewExit. 854 PHINode *PN; 855 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) { 856 PN = cast<PHINode>(I); 857 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 858 ValueToValueMapTy::iterator It = VMap.find(V); 859 if (It != VMap.end()) V = It->second; 860 PN->addIncoming(V, NewExit); 861 } 862 863 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) { 864 PN = PHINode::Create(LPad->getType(), 0, "", 865 ExitSucc->getFirstInsertionPt()); 866 867 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc); 868 I != E; ++I) { 869 BasicBlock *BB = *I; 870 LandingPadInst *LPI = BB->getLandingPadInst(); 871 LPI->replaceAllUsesWith(PN); 872 PN->addIncoming(LPI, BB); 873 } 874 } 875 } 876 877 // Rewrite the code to refer to itself. 878 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 879 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 880 E = NewBlocks[i]->end(); I != E; ++I) 881 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); 882 883 // Rewrite the original preheader to select between versions of the loop. 884 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 885 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 886 "Preheader splitting did not work correctly!"); 887 888 // Emit the new branch that selects between the two versions of this loop. 889 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 890 LPM->deleteSimpleAnalysisValue(OldBR, L); 891 OldBR->eraseFromParent(); 892 893 LoopProcessWorklist.push_back(NewLoop); 894 redoLoop = true; 895 896 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody 897 // deletes the instruction (for example by simplifying a PHI that feeds into 898 // the condition that we're unswitching on), we don't rewrite the second 899 // iteration. 900 WeakVH LICHandle(LIC); 901 902 // Now we rewrite the original code to know that the condition is true and the 903 // new code to know that the condition is false. 904 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); 905 906 // It's possible that simplifying one loop could cause the other to be 907 // changed to another value or a constant. If its a constant, don't simplify 908 // it. 909 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop && 910 LICHandle && !isa<Constant>(LICHandle)) 911 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); 912} 913 914/// RemoveFromWorklist - Remove all instances of I from the worklist vector 915/// specified. 916static void RemoveFromWorklist(Instruction *I, 917 std::vector<Instruction*> &Worklist) { 918 919 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I), 920 Worklist.end()); 921} 922 923/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 924/// program, replacing all uses with V and update the worklist. 925static void ReplaceUsesOfWith(Instruction *I, Value *V, 926 std::vector<Instruction*> &Worklist, 927 Loop *L, LPPassManager *LPM) { 928 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 929 930 // Add uses to the worklist, which may be dead now. 931 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 932 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 933 Worklist.push_back(Use); 934 935 // Add users to the worklist which may be simplified now. 936 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 937 UI != E; ++UI) 938 Worklist.push_back(cast<Instruction>(*UI)); 939 LPM->deleteSimpleAnalysisValue(I, L); 940 RemoveFromWorklist(I, Worklist); 941 I->replaceAllUsesWith(V); 942 I->eraseFromParent(); 943 ++NumSimplify; 944} 945 946/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 947/// information, and remove any dead successors it has. 948/// 949void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 950 std::vector<Instruction*> &Worklist, 951 Loop *L) { 952 if (pred_begin(BB) != pred_end(BB)) { 953 // This block isn't dead, since an edge to BB was just removed, see if there 954 // are any easy simplifications we can do now. 955 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 956 // If it has one pred, fold phi nodes in BB. 957 while (isa<PHINode>(BB->begin())) 958 ReplaceUsesOfWith(BB->begin(), 959 cast<PHINode>(BB->begin())->getIncomingValue(0), 960 Worklist, L, LPM); 961 962 // If this is the header of a loop and the only pred is the latch, we now 963 // have an unreachable loop. 964 if (Loop *L = LI->getLoopFor(BB)) 965 if (loopHeader == BB && L->contains(Pred)) { 966 // Remove the branch from the latch to the header block, this makes 967 // the header dead, which will make the latch dead (because the header 968 // dominates the latch). 969 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 970 Pred->getTerminator()->eraseFromParent(); 971 new UnreachableInst(BB->getContext(), Pred); 972 973 // The loop is now broken, remove it from LI. 974 RemoveLoopFromHierarchy(L); 975 976 // Reprocess the header, which now IS dead. 977 RemoveBlockIfDead(BB, Worklist, L); 978 return; 979 } 980 981 // If pred ends in a uncond branch, add uncond branch to worklist so that 982 // the two blocks will get merged. 983 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 984 if (BI->isUnconditional()) 985 Worklist.push_back(BI); 986 } 987 return; 988 } 989 990 DEBUG(dbgs() << "Nuking dead block: " << *BB); 991 992 // Remove the instructions in the basic block from the worklist. 993 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 994 RemoveFromWorklist(I, Worklist); 995 996 // Anything that uses the instructions in this basic block should have their 997 // uses replaced with undefs. 998 // If I is not void type then replaceAllUsesWith undef. 999 // This allows ValueHandlers and custom metadata to adjust itself. 1000 if (!I->getType()->isVoidTy()) 1001 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1002 } 1003 1004 // If this is the edge to the header block for a loop, remove the loop and 1005 // promote all subloops. 1006 if (Loop *BBLoop = LI->getLoopFor(BB)) { 1007 if (BBLoop->getLoopLatch() == BB) { 1008 RemoveLoopFromHierarchy(BBLoop); 1009 if (currentLoop == BBLoop) { 1010 currentLoop = 0; 1011 redoLoop = false; 1012 } 1013 } 1014 } 1015 1016 // Remove the block from the loop info, which removes it from any loops it 1017 // was in. 1018 LI->removeBlock(BB); 1019 1020 1021 // Remove phi node entries in successors for this block. 1022 TerminatorInst *TI = BB->getTerminator(); 1023 SmallVector<BasicBlock*, 4> Succs; 1024 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1025 Succs.push_back(TI->getSuccessor(i)); 1026 TI->getSuccessor(i)->removePredecessor(BB); 1027 } 1028 1029 // Unique the successors, remove anything with multiple uses. 1030 array_pod_sort(Succs.begin(), Succs.end()); 1031 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 1032 1033 // Remove the basic block, including all of the instructions contained in it. 1034 LPM->deleteSimpleAnalysisValue(BB, L); 1035 BB->eraseFromParent(); 1036 // Remove successor blocks here that are not dead, so that we know we only 1037 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 1038 // then getting removed before we revisit them, which is badness. 1039 // 1040 for (unsigned i = 0; i != Succs.size(); ++i) 1041 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 1042 // One exception is loop headers. If this block was the preheader for a 1043 // loop, then we DO want to visit the loop so the loop gets deleted. 1044 // We know that if the successor is a loop header, that this loop had to 1045 // be the preheader: the case where this was the latch block was handled 1046 // above and headers can only have two predecessors. 1047 if (!LI->isLoopHeader(Succs[i])) { 1048 Succs.erase(Succs.begin()+i); 1049 --i; 1050 } 1051 } 1052 1053 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 1054 RemoveBlockIfDead(Succs[i], Worklist, L); 1055} 1056 1057/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 1058/// become unwrapped, either because the backedge was deleted, or because the 1059/// edge into the header was removed. If the edge into the header from the 1060/// latch block was removed, the loop is unwrapped but subloops are still alive, 1061/// so they just reparent loops. If the loops are actually dead, they will be 1062/// removed later. 1063void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 1064 LPM->deleteLoopFromQueue(L); 1065 RemoveLoopFromWorklist(L); 1066} 1067 1068// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 1069// the value specified by Val in the specified loop, or we know it does NOT have 1070// that value. Rewrite any uses of LIC or of properties correlated to it. 1071void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 1072 Constant *Val, 1073 bool IsEqual) { 1074 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 1075 1076 // FIXME: Support correlated properties, like: 1077 // for (...) 1078 // if (li1 < li2) 1079 // ... 1080 // if (li1 > li2) 1081 // ... 1082 1083 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 1084 // selects, switches. 1085 std::vector<Instruction*> Worklist; 1086 LLVMContext &Context = Val->getContext(); 1087 1088 1089 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 1090 // in the loop with the appropriate one directly. 1091 if (IsEqual || (isa<ConstantInt>(Val) && 1092 Val->getType()->isIntegerTy(1))) { 1093 Value *Replacement; 1094 if (IsEqual) 1095 Replacement = Val; 1096 else 1097 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 1098 !cast<ConstantInt>(Val)->getZExtValue()); 1099 1100 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); 1101 UI != E; ++UI) { 1102 Instruction *U = dyn_cast<Instruction>(*UI); 1103 if (!U || !L->contains(U)) 1104 continue; 1105 Worklist.push_back(U); 1106 } 1107 1108 for (std::vector<Instruction*>::iterator UI = Worklist.begin(); 1109 UI != Worklist.end(); ++UI) 1110 (*UI)->replaceUsesOfWith(LIC, Replacement); 1111 1112 SimplifyCode(Worklist, L); 1113 return; 1114 } 1115 1116 // Otherwise, we don't know the precise value of LIC, but we do know that it 1117 // is certainly NOT "Val". As such, simplify any uses in the loop that we 1118 // can. This case occurs when we unswitch switch statements. 1119 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); 1120 UI != E; ++UI) { 1121 Instruction *U = dyn_cast<Instruction>(*UI); 1122 if (!U || !L->contains(U)) 1123 continue; 1124 1125 Worklist.push_back(U); 1126 1127 // TODO: We could do other simplifications, for example, turning 1128 // 'icmp eq LIC, Val' -> false. 1129 1130 // If we know that LIC is not Val, use this info to simplify code. 1131 SwitchInst *SI = dyn_cast<SwitchInst>(U); 1132 if (SI == 0 || !isa<ConstantInt>(Val)) continue; 1133 1134 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val)); 1135 // Default case is live for multiple values. 1136 if (DeadCase == SI->case_default()) continue; 1137 1138 // Found a dead case value. Don't remove PHI nodes in the 1139 // successor if they become single-entry, those PHI nodes may 1140 // be in the Users list. 1141 1142 BasicBlock *Switch = SI->getParent(); 1143 BasicBlock *SISucc = DeadCase.getCaseSuccessor(); 1144 BasicBlock *Latch = L->getLoopLatch(); 1145 1146 BranchesInfo.setUnswitched(SI, Val); 1147 1148 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical. 1149 // If the DeadCase successor dominates the loop latch, then the 1150 // transformation isn't safe since it will delete the sole predecessor edge 1151 // to the latch. 1152 if (Latch && DT->dominates(SISucc, Latch)) 1153 continue; 1154 1155 // FIXME: This is a hack. We need to keep the successor around 1156 // and hooked up so as to preserve the loop structure, because 1157 // trying to update it is complicated. So instead we preserve the 1158 // loop structure and put the block on a dead code path. 1159 SplitEdge(Switch, SISucc, this); 1160 // Compute the successors instead of relying on the return value 1161 // of SplitEdge, since it may have split the switch successor 1162 // after PHI nodes. 1163 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor(); 1164 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 1165 // Create an "unreachable" destination. 1166 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 1167 Switch->getParent(), 1168 OldSISucc); 1169 new UnreachableInst(Context, Abort); 1170 // Force the new case destination to branch to the "unreachable" 1171 // block while maintaining a (dead) CFG edge to the old block. 1172 NewSISucc->getTerminator()->eraseFromParent(); 1173 BranchInst::Create(Abort, OldSISucc, 1174 ConstantInt::getTrue(Context), NewSISucc); 1175 // Release the PHI operands for this edge. 1176 for (BasicBlock::iterator II = NewSISucc->begin(); 1177 PHINode *PN = dyn_cast<PHINode>(II); ++II) 1178 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 1179 UndefValue::get(PN->getType())); 1180 // Tell the domtree about the new block. We don't fully update the 1181 // domtree here -- instead we force it to do a full recomputation 1182 // after the pass is complete -- but we do need to inform it of 1183 // new blocks. 1184 if (DT) 1185 DT->addNewBlock(Abort, NewSISucc); 1186 } 1187 1188 SimplifyCode(Worklist, L); 1189} 1190 1191/// SimplifyCode - Okay, now that we have simplified some instructions in the 1192/// loop, walk over it and constant prop, dce, and fold control flow where 1193/// possible. Note that this is effectively a very simple loop-structure-aware 1194/// optimizer. During processing of this loop, L could very well be deleted, so 1195/// it must not be used. 1196/// 1197/// FIXME: When the loop optimizer is more mature, separate this out to a new 1198/// pass. 1199/// 1200void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 1201 while (!Worklist.empty()) { 1202 Instruction *I = Worklist.back(); 1203 Worklist.pop_back(); 1204 1205 // Simple DCE. 1206 if (isInstructionTriviallyDead(I)) { 1207 DEBUG(dbgs() << "Remove dead instruction '" << *I); 1208 1209 // Add uses to the worklist, which may be dead now. 1210 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1211 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1212 Worklist.push_back(Use); 1213 LPM->deleteSimpleAnalysisValue(I, L); 1214 RemoveFromWorklist(I, Worklist); 1215 I->eraseFromParent(); 1216 ++NumSimplify; 1217 continue; 1218 } 1219 1220 // See if instruction simplification can hack this up. This is common for 1221 // things like "select false, X, Y" after unswitching made the condition be 1222 // 'false'. TODO: update the domtree properly so we can pass it here. 1223 if (Value *V = SimplifyInstruction(I)) 1224 if (LI->replacementPreservesLCSSAForm(I, V)) { 1225 ReplaceUsesOfWith(I, V, Worklist, L, LPM); 1226 continue; 1227 } 1228 1229 // Special case hacks that appear commonly in unswitched code. 1230 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1231 if (BI->isUnconditional()) { 1232 // If BI's parent is the only pred of the successor, fold the two blocks 1233 // together. 1234 BasicBlock *Pred = BI->getParent(); 1235 BasicBlock *Succ = BI->getSuccessor(0); 1236 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1237 if (!SinglePred) continue; // Nothing to do. 1238 assert(SinglePred == Pred && "CFG broken"); 1239 1240 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1241 << Succ->getName() << "\n"); 1242 1243 // Resolve any single entry PHI nodes in Succ. 1244 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1245 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1246 1247 // If Succ has any successors with PHI nodes, update them to have 1248 // entries coming from Pred instead of Succ. 1249 Succ->replaceAllUsesWith(Pred); 1250 1251 // Move all of the successor contents from Succ to Pred. 1252 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1253 Succ->end()); 1254 LPM->deleteSimpleAnalysisValue(BI, L); 1255 BI->eraseFromParent(); 1256 RemoveFromWorklist(BI, Worklist); 1257 1258 // Remove Succ from the loop tree. 1259 LI->removeBlock(Succ); 1260 LPM->deleteSimpleAnalysisValue(Succ, L); 1261 Succ->eraseFromParent(); 1262 ++NumSimplify; 1263 continue; 1264 } 1265 1266 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1267 // Conditional branch. Turn it into an unconditional branch, then 1268 // remove dead blocks. 1269 continue; // FIXME: Enable. 1270 1271 DEBUG(dbgs() << "Folded branch: " << *BI); 1272 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1273 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1274 DeadSucc->removePredecessor(BI->getParent(), true); 1275 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1276 LPM->deleteSimpleAnalysisValue(BI, L); 1277 BI->eraseFromParent(); 1278 RemoveFromWorklist(BI, Worklist); 1279 ++NumSimplify; 1280 1281 RemoveBlockIfDead(DeadSucc, Worklist, L); 1282 } 1283 continue; 1284 } 1285 } 1286} 1287