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/Constants.h" 32 #include "llvm/DerivedTypes.h" 33 #include "llvm/Function.h" 34 #include "llvm/Instructions.h" 35 #include "llvm/Analysis/CodeMetrics.h" 36 #include "llvm/Analysis/InstructionSimplify.h" 37 #include "llvm/Analysis/LoopInfo.h" 38 #include "llvm/Analysis/LoopPass.h" 39 #include "llvm/Analysis/Dominators.h" 40 #include "llvm/Analysis/ScalarEvolution.h" 41 #include "llvm/Transforms/Utils/Cloning.h" 42 #include "llvm/Transforms/Utils/Local.h" 43 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 44 #include "llvm/ADT/Statistic.h" 45 #include "llvm/ADT/SmallPtrSet.h" 46 #include "llvm/ADT/STLExtras.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/Debug.h" 49 #include "llvm/Support/raw_ostream.h" 50 #include <algorithm> 51 #include <map> 52 #include <set> 53 using namespace llvm; 54 55 STATISTIC(NumBranches, "Number of branches unswitched"); 56 STATISTIC(NumSwitches, "Number of switches unswitched"); 57 STATISTIC(NumSelects , "Number of selects unswitched"); 58 STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 59 STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 60 STATISTIC(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. 64 static cl::opt<unsigned> 65 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 66 cl::init(100), cl::Hidden); 67 68 namespace { 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. 224 bool 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 } 252 253 if (!Props.CanBeUnswitchedCount) { 254 DEBUG(dbgs() << "NOT unswitching loop %" 255 << L->getHeader()->getName() << ", cost too high: " 256 << L->getBlocks().size() << "\n"); 257 258 return false; 259 } 260 261 // Be careful. This links are good only before new loop addition. 262 CurrentLoopProperties = &Props; 263 CurLoopInstructions = &Props.UnswitchedVals; 264 265 return true; 266 } 267 268 // Clean all data related to given loop. 269 void LUAnalysisCache::forgetLoop(const Loop* L) { 270 271 LoopPropsMapIt LIt = LoopsProperties.find(L); 272 273 if (LIt != LoopsProperties.end()) { 274 LoopProperties& Props = LIt->second; 275 MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation; 276 LoopsProperties.erase(LIt); 277 } 278 279 CurrentLoopProperties = NULL; 280 CurLoopInstructions = NULL; 281 } 282 283 // Mark case value as unswitched. 284 // Since SI instruction can be partly unswitched, in order to avoid 285 // extra unswitching in cloned loops keep track all unswitched values. 286 void LUAnalysisCache::setUnswitched(const SwitchInst* SI, const Value* V) { 287 (*CurLoopInstructions)[SI].insert(V); 288 } 289 290 // Check was this case value unswitched before or not. 291 bool LUAnalysisCache::isUnswitched(const SwitchInst* SI, const Value* V) { 292 return (*CurLoopInstructions)[SI].count(V); 293 } 294 295 // Clone all loop-unswitch related loop properties. 296 // Redistribute unswitching quotas. 297 // Note, that new loop data is stored inside the VMap. 298 void LUAnalysisCache::cloneData(const Loop* NewLoop, const Loop* OldLoop, 299 const ValueToValueMapTy& VMap) { 300 301 LoopProperties& NewLoopProps = LoopsProperties[NewLoop]; 302 LoopProperties& OldLoopProps = *CurrentLoopProperties; 303 UnswitchedValsMap& Insts = OldLoopProps.UnswitchedVals; 304 305 // Reallocate "can-be-unswitched quota" 306 307 --OldLoopProps.CanBeUnswitchedCount; 308 unsigned Quota = OldLoopProps.CanBeUnswitchedCount; 309 NewLoopProps.CanBeUnswitchedCount = Quota / 2; 310 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2; 311 312 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation; 313 314 // Clone unswitched values info: 315 // for new loop switches we clone info about values that was 316 // already unswitched and has redundant successors. 317 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) { 318 const SwitchInst* OldInst = I->first; 319 Value* NewI = VMap.lookup(OldInst); 320 const SwitchInst* NewInst = cast_or_null<SwitchInst>(NewI); 321 assert(NewInst && "All instructions that are in SrcBB must be in VMap."); 322 323 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst]; 324 } 325 } 326 327 char LoopUnswitch::ID = 0; 328 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", 329 false, false) 330 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 331 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 332 INITIALIZE_PASS_DEPENDENCY(LCSSA) 333 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", 334 false, false) 335 336 Pass *llvm::createLoopUnswitchPass(bool Os) { 337 return new LoopUnswitch(Os); 338 } 339 340 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 341 /// invariant in the loop, or has an invariant piece, return the invariant. 342 /// Otherwise, return null. 343 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 344 345 // We started analyze new instruction, increment scanned instructions counter. 346 ++TotalInsts; 347 348 // We can never unswitch on vector conditions. 349 if (Cond->getType()->isVectorTy()) 350 return 0; 351 352 // Constants should be folded, not unswitched on! 353 if (isa<Constant>(Cond)) return 0; 354 355 // TODO: Handle: br (VARIANT|INVARIANT). 356 357 // Hoist simple values out. 358 if (L->makeLoopInvariant(Cond, Changed)) 359 return Cond; 360 361 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 362 if (BO->getOpcode() == Instruction::And || 363 BO->getOpcode() == Instruction::Or) { 364 // If either the left or right side is invariant, we can unswitch on this, 365 // which will cause the branch to go away in one loop and the condition to 366 // simplify in the other one. 367 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 368 return LHS; 369 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 370 return RHS; 371 } 372 373 return 0; 374 } 375 376 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 377 LI = &getAnalysis<LoopInfo>(); 378 LPM = &LPM_Ref; 379 DT = getAnalysisIfAvailable<DominatorTree>(); 380 currentLoop = L; 381 Function *F = currentLoop->getHeader()->getParent(); 382 bool Changed = false; 383 do { 384 assert(currentLoop->isLCSSAForm(*DT)); 385 redoLoop = false; 386 Changed |= processCurrentLoop(); 387 } while(redoLoop); 388 389 if (Changed) { 390 // FIXME: Reconstruct dom info, because it is not preserved properly. 391 if (DT) 392 DT->runOnFunction(*F); 393 } 394 return Changed; 395 } 396 397 /// processCurrentLoop - Do actual work and unswitch loop if possible 398 /// and profitable. 399 bool LoopUnswitch::processCurrentLoop() { 400 bool Changed = false; 401 402 initLoopData(); 403 404 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 405 if (!loopPreheader) 406 return false; 407 408 // Loops with indirectbr cannot be cloned. 409 if (!currentLoop->isSafeToClone()) 410 return false; 411 412 // Loops with invokes, whose unwind edge escapes the loop, cannot be 413 // unswitched because splitting their edges are non-trivial and don't preserve 414 // loop simplify information. 415 for (Loop::block_iterator I = currentLoop->block_begin(), 416 E = currentLoop->block_end(); I != E; ++I) 417 if (const InvokeInst *II = dyn_cast<InvokeInst>((*I)->getTerminator())) 418 if (!currentLoop->contains(II->getUnwindDest())) 419 return false; 420 421 // Without dedicated exits, splitting the exit edge may fail. 422 if (!currentLoop->hasDedicatedExits()) 423 return false; 424 425 LLVMContext &Context = loopHeader->getContext(); 426 427 // Probably we reach the quota of branches for this loop. If so 428 // stop unswitching. 429 if (!BranchesInfo.countLoop(currentLoop)) 430 return false; 431 432 // Loop over all of the basic blocks in the loop. If we find an interior 433 // block that is branching on a loop-invariant condition, we can unswitch this 434 // loop. 435 for (Loop::block_iterator I = currentLoop->block_begin(), 436 E = currentLoop->block_end(); I != E; ++I) { 437 TerminatorInst *TI = (*I)->getTerminator(); 438 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 439 // If this isn't branching on an invariant condition, we can't unswitch 440 // it. 441 if (BI->isConditional()) { 442 // See if this, or some part of it, is loop invariant. If so, we can 443 // unswitch on it if we desire. 444 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 445 currentLoop, Changed); 446 if (LoopCond && UnswitchIfProfitable(LoopCond, 447 ConstantInt::getTrue(Context))) { 448 ++NumBranches; 449 return true; 450 } 451 } 452 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 453 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 454 currentLoop, Changed); 455 unsigned NumCases = SI->getNumCases(); 456 if (LoopCond && NumCases) { 457 // Find a value to unswitch on: 458 // FIXME: this should chose the most expensive case! 459 // FIXME: scan for a case with a non-critical edge? 460 Constant *UnswitchVal = NULL; 461 462 // Do not process same value again and again. 463 // At this point we have some cases already unswitched and 464 // some not yet unswitched. Let's find the first not yet unswitched one. 465 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 466 i != e; ++i) { 467 Constant* UnswitchValCandidate = i.getCaseValue(); 468 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) { 469 UnswitchVal = UnswitchValCandidate; 470 break; 471 } 472 } 473 474 if (!UnswitchVal) 475 continue; 476 477 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 478 ++NumSwitches; 479 return true; 480 } 481 } 482 } 483 484 // Scan the instructions to check for unswitchable values. 485 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 486 BBI != E; ++BBI) 487 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 488 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 489 currentLoop, Changed); 490 if (LoopCond && UnswitchIfProfitable(LoopCond, 491 ConstantInt::getTrue(Context))) { 492 ++NumSelects; 493 return true; 494 } 495 } 496 } 497 return Changed; 498 } 499 500 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the 501 /// loop with no side effects (including infinite loops). 502 /// 503 /// If true, we return true and set ExitBB to the block we 504 /// exit through. 505 /// 506 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 507 BasicBlock *&ExitBB, 508 std::set<BasicBlock*> &Visited) { 509 if (!Visited.insert(BB).second) { 510 // Already visited. Without more analysis, this could indicate an infinite 511 // loop. 512 return false; 513 } else if (!L->contains(BB)) { 514 // Otherwise, this is a loop exit, this is fine so long as this is the 515 // first exit. 516 if (ExitBB != 0) return false; 517 ExitBB = BB; 518 return true; 519 } 520 521 // Otherwise, this is an unvisited intra-loop node. Check all successors. 522 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 523 // Check to see if the successor is a trivial loop exit. 524 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 525 return false; 526 } 527 528 // Okay, everything after this looks good, check to make sure that this block 529 // doesn't include any side effects. 530 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 531 if (I->mayHaveSideEffects()) 532 return false; 533 534 return true; 535 } 536 537 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally 538 /// leads to an exit from the specified loop, and has no side-effects in the 539 /// process. If so, return the block that is exited to, otherwise return null. 540 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 541 std::set<BasicBlock*> Visited; 542 Visited.insert(L->getHeader()); // Branches to header make infinite loops. 543 BasicBlock *ExitBB = 0; 544 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 545 return ExitBB; 546 return 0; 547 } 548 549 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 550 /// trivial: that is, that the condition controls whether or not the loop does 551 /// anything at all. If this is a trivial condition, unswitching produces no 552 /// code duplications (equivalently, it produces a simpler loop and a new empty 553 /// loop, which gets deleted). 554 /// 555 /// If this is a trivial condition, return true, otherwise return false. When 556 /// returning true, this sets Cond and Val to the condition that controls the 557 /// trivial condition: when Cond dynamically equals Val, the loop is known to 558 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when 559 /// Cond == Val. 560 /// 561 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 562 BasicBlock **LoopExit) { 563 BasicBlock *Header = currentLoop->getHeader(); 564 TerminatorInst *HeaderTerm = Header->getTerminator(); 565 LLVMContext &Context = Header->getContext(); 566 567 BasicBlock *LoopExitBB = 0; 568 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 569 // If the header block doesn't end with a conditional branch on Cond, we 570 // can't handle it. 571 if (!BI->isConditional() || BI->getCondition() != Cond) 572 return false; 573 574 // Check to see if a successor of the branch is guaranteed to 575 // exit through a unique exit block without having any 576 // side-effects. If so, determine the value of Cond that causes it to do 577 // this. 578 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 579 BI->getSuccessor(0)))) { 580 if (Val) *Val = ConstantInt::getTrue(Context); 581 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 582 BI->getSuccessor(1)))) { 583 if (Val) *Val = ConstantInt::getFalse(Context); 584 } 585 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 586 // If this isn't a switch on Cond, we can't handle it. 587 if (SI->getCondition() != Cond) return false; 588 589 // Check to see if a successor of the switch is guaranteed to go to the 590 // latch block or exit through a one exit block without having any 591 // side-effects. If so, determine the value of Cond that causes it to do 592 // this. 593 // Note that we can't trivially unswitch on the default case or 594 // on already unswitched cases. 595 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 596 i != e; ++i) { 597 BasicBlock* LoopExitCandidate; 598 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop, 599 i.getCaseSuccessor()))) { 600 // Okay, we found a trivial case, remember the value that is trivial. 601 ConstantInt* CaseVal = i.getCaseValue(); 602 603 // Check that it was not unswitched before, since already unswitched 604 // trivial vals are looks trivial too. 605 if (BranchesInfo.isUnswitched(SI, CaseVal)) 606 continue; 607 LoopExitBB = LoopExitCandidate; 608 if (Val) *Val = CaseVal; 609 break; 610 } 611 } 612 } 613 614 // If we didn't find a single unique LoopExit block, or if the loop exit block 615 // contains phi nodes, this isn't trivial. 616 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 617 return false; // Can't handle this. 618 619 if (LoopExit) *LoopExit = LoopExitBB; 620 621 // We already know that nothing uses any scalar values defined inside of this 622 // loop. As such, we just have to check to see if this loop will execute any 623 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 624 // part of the loop that the code *would* execute. We already checked the 625 // tail, check the header now. 626 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 627 if (I->mayHaveSideEffects()) 628 return false; 629 return true; 630 } 631 632 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 633 /// LoopCond == Val to simplify the loop. If we decide that this is profitable, 634 /// unswitch the loop, reprocess the pieces, then return true. 635 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { 636 637 Function *F = loopHeader->getParent(); 638 639 Constant *CondVal = 0; 640 BasicBlock *ExitBlock = 0; 641 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 642 // If the condition is trivial, always unswitch. There is no code growth 643 // for this case. 644 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 645 return true; 646 } 647 648 // Check to see if it would be profitable to unswitch current loop. 649 650 // Do not do non-trivial unswitch while optimizing for size. 651 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) 652 return false; 653 654 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 655 return true; 656 } 657 658 /// CloneLoop - Recursively clone the specified loop and all of its children, 659 /// mapping the blocks with the specified map. 660 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, 661 LoopInfo *LI, LPPassManager *LPM) { 662 Loop *New = new Loop(); 663 LPM->insertLoop(New, PL); 664 665 // Add all of the blocks in L to the new loop. 666 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 667 I != E; ++I) 668 if (LI->getLoopFor(*I) == L) 669 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 670 671 // Add all of the subloops to the new loop. 672 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 673 CloneLoop(*I, New, VM, LI, LPM); 674 675 return New; 676 } 677 678 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 679 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 680 /// code immediately before InsertPt. 681 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 682 BasicBlock *TrueDest, 683 BasicBlock *FalseDest, 684 Instruction *InsertPt) { 685 // Insert a conditional branch on LIC to the two preheaders. The original 686 // code is the true version and the new code is the false version. 687 Value *BranchVal = LIC; 688 if (!isa<ConstantInt>(Val) || 689 Val->getType() != Type::getInt1Ty(LIC->getContext())) 690 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val); 691 else if (Val != ConstantInt::getTrue(Val->getContext())) 692 // We want to enter the new loop when the condition is true. 693 std::swap(TrueDest, FalseDest); 694 695 // Insert the new branch. 696 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 697 698 // If either edge is critical, split it. This helps preserve LoopSimplify 699 // form for enclosing loops. 700 SplitCriticalEdge(BI, 0, this); 701 SplitCriticalEdge(BI, 1, this); 702 } 703 704 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 705 /// condition in it (a cond branch from its header block to its latch block, 706 /// where the path through the loop that doesn't execute its body has no 707 /// side-effects), unswitch it. This doesn't involve any code duplication, just 708 /// moving the conditional branch outside of the loop and updating loop info. 709 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 710 Constant *Val, 711 BasicBlock *ExitBlock) { 712 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 713 << loopHeader->getName() << " [" << L->getBlocks().size() 714 << " blocks] in Function " << L->getHeader()->getParent()->getName() 715 << " on cond: " << *Val << " == " << *Cond << "\n"); 716 717 // First step, split the preheader, so that we know that there is a safe place 718 // to insert the conditional branch. We will change loopPreheader to have a 719 // conditional branch on Cond. 720 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 721 722 // Now that we have a place to insert the conditional branch, create a place 723 // to branch to: this is the exit block out of the loop that we should 724 // short-circuit to. 725 726 // Split this block now, so that the loop maintains its exit block, and so 727 // that the jump from the preheader can execute the contents of the exit block 728 // without actually branching to it (the exit block should be dominated by the 729 // loop header, not the preheader). 730 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 731 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 732 733 // Okay, now we have a position to branch from and a position to branch to, 734 // insert the new conditional branch. 735 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 736 loopPreheader->getTerminator()); 737 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 738 loopPreheader->getTerminator()->eraseFromParent(); 739 740 // We need to reprocess this loop, it could be unswitched again. 741 redoLoop = true; 742 743 // Now that we know that the loop is never entered when this condition is a 744 // particular value, rewrite the loop with this info. We know that this will 745 // at least eliminate the old branch. 746 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 747 ++NumTrivial; 748 } 749 750 /// SplitExitEdges - Split all of the edges from inside the loop to their exit 751 /// blocks. Update the appropriate Phi nodes as we do so. 752 void LoopUnswitch::SplitExitEdges(Loop *L, 753 const SmallVector<BasicBlock *, 8> &ExitBlocks){ 754 755 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 756 BasicBlock *ExitBlock = ExitBlocks[i]; 757 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 758 pred_end(ExitBlock)); 759 760 // Although SplitBlockPredecessors doesn't preserve loop-simplify in 761 // general, if we call it on all predecessors of all exits then it does. 762 if (!ExitBlock->isLandingPad()) { 763 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this); 764 } else { 765 SmallVector<BasicBlock*, 2> NewBBs; 766 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa", 767 this, NewBBs); 768 } 769 } 770 } 771 772 /// UnswitchNontrivialCondition - We determined that the loop is profitable 773 /// to unswitch when LIC equal Val. Split it into loop versions and test the 774 /// condition outside of either loop. Return the loops created as Out1/Out2. 775 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 776 Loop *L) { 777 Function *F = loopHeader->getParent(); 778 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 779 << loopHeader->getName() << " [" << L->getBlocks().size() 780 << " blocks] in Function " << F->getName() 781 << " when '" << *Val << "' == " << *LIC << "\n"); 782 783 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>()) 784 SE->forgetLoop(L); 785 786 LoopBlocks.clear(); 787 NewBlocks.clear(); 788 789 // First step, split the preheader and exit blocks, and add these blocks to 790 // the LoopBlocks list. 791 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 792 LoopBlocks.push_back(NewPreheader); 793 794 // We want the loop to come after the preheader, but before the exit blocks. 795 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 796 797 SmallVector<BasicBlock*, 8> ExitBlocks; 798 L->getUniqueExitBlocks(ExitBlocks); 799 800 // Split all of the edges from inside the loop to their exit blocks. Update 801 // the appropriate Phi nodes as we do so. 802 SplitExitEdges(L, ExitBlocks); 803 804 // The exit blocks may have been changed due to edge splitting, recompute. 805 ExitBlocks.clear(); 806 L->getUniqueExitBlocks(ExitBlocks); 807 808 // Add exit blocks to the loop blocks. 809 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 810 811 // Next step, clone all of the basic blocks that make up the loop (including 812 // the loop preheader and exit blocks), keeping track of the mapping between 813 // the instructions and blocks. 814 NewBlocks.reserve(LoopBlocks.size()); 815 ValueToValueMapTy VMap; 816 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 817 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); 818 819 NewBlocks.push_back(NewBB); 820 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. 821 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); 822 } 823 824 // Splice the newly inserted blocks into the function right before the 825 // original preheader. 826 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(), 827 NewBlocks[0], F->end()); 828 829 // Now we create the new Loop object for the versioned loop. 830 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); 831 832 // Recalculate unswitching quota, inherit simplified switches info for NewBB, 833 // Probably clone more loop-unswitch related loop properties. 834 BranchesInfo.cloneData(NewLoop, L, VMap); 835 836 Loop *ParentLoop = L->getParentLoop(); 837 if (ParentLoop) { 838 // Make sure to add the cloned preheader and exit blocks to the parent loop 839 // as well. 840 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 841 } 842 843 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 844 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]); 845 // The new exit block should be in the same loop as the old one. 846 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 847 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 848 849 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 850 "Exit block should have been split to have one successor!"); 851 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 852 853 // If the successor of the exit block had PHI nodes, add an entry for 854 // NewExit. 855 PHINode *PN; 856 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) { 857 PN = cast<PHINode>(I); 858 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 859 ValueToValueMapTy::iterator It = VMap.find(V); 860 if (It != VMap.end()) V = It->second; 861 PN->addIncoming(V, NewExit); 862 } 863 864 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) { 865 PN = PHINode::Create(LPad->getType(), 0, "", 866 ExitSucc->getFirstInsertionPt()); 867 868 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc); 869 I != E; ++I) { 870 BasicBlock *BB = *I; 871 LandingPadInst *LPI = BB->getLandingPadInst(); 872 LPI->replaceAllUsesWith(PN); 873 PN->addIncoming(LPI, BB); 874 } 875 } 876 } 877 878 // Rewrite the code to refer to itself. 879 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 880 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 881 E = NewBlocks[i]->end(); I != E; ++I) 882 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); 883 884 // Rewrite the original preheader to select between versions of the loop. 885 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 886 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 887 "Preheader splitting did not work correctly!"); 888 889 // Emit the new branch that selects between the two versions of this loop. 890 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 891 LPM->deleteSimpleAnalysisValue(OldBR, L); 892 OldBR->eraseFromParent(); 893 894 LoopProcessWorklist.push_back(NewLoop); 895 redoLoop = true; 896 897 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody 898 // deletes the instruction (for example by simplifying a PHI that feeds into 899 // the condition that we're unswitching on), we don't rewrite the second 900 // iteration. 901 WeakVH LICHandle(LIC); 902 903 // Now we rewrite the original code to know that the condition is true and the 904 // new code to know that the condition is false. 905 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); 906 907 // It's possible that simplifying one loop could cause the other to be 908 // changed to another value or a constant. If its a constant, don't simplify 909 // it. 910 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop && 911 LICHandle && !isa<Constant>(LICHandle)) 912 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); 913 } 914 915 /// RemoveFromWorklist - Remove all instances of I from the worklist vector 916 /// specified. 917 static void RemoveFromWorklist(Instruction *I, 918 std::vector<Instruction*> &Worklist) { 919 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 920 Worklist.end(), I); 921 while (WI != Worklist.end()) { 922 unsigned Offset = WI-Worklist.begin(); 923 Worklist.erase(WI); 924 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 925 } 926 } 927 928 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 929 /// program, replacing all uses with V and update the worklist. 930 static void ReplaceUsesOfWith(Instruction *I, Value *V, 931 std::vector<Instruction*> &Worklist, 932 Loop *L, LPPassManager *LPM) { 933 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 934 935 // Add uses to the worklist, which may be dead now. 936 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 937 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 938 Worklist.push_back(Use); 939 940 // Add users to the worklist which may be simplified now. 941 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 942 UI != E; ++UI) 943 Worklist.push_back(cast<Instruction>(*UI)); 944 LPM->deleteSimpleAnalysisValue(I, L); 945 RemoveFromWorklist(I, Worklist); 946 I->replaceAllUsesWith(V); 947 I->eraseFromParent(); 948 ++NumSimplify; 949 } 950 951 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 952 /// information, and remove any dead successors it has. 953 /// 954 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 955 std::vector<Instruction*> &Worklist, 956 Loop *L) { 957 if (pred_begin(BB) != pred_end(BB)) { 958 // This block isn't dead, since an edge to BB was just removed, see if there 959 // are any easy simplifications we can do now. 960 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 961 // If it has one pred, fold phi nodes in BB. 962 while (isa<PHINode>(BB->begin())) 963 ReplaceUsesOfWith(BB->begin(), 964 cast<PHINode>(BB->begin())->getIncomingValue(0), 965 Worklist, L, LPM); 966 967 // If this is the header of a loop and the only pred is the latch, we now 968 // have an unreachable loop. 969 if (Loop *L = LI->getLoopFor(BB)) 970 if (loopHeader == BB && L->contains(Pred)) { 971 // Remove the branch from the latch to the header block, this makes 972 // the header dead, which will make the latch dead (because the header 973 // dominates the latch). 974 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 975 Pred->getTerminator()->eraseFromParent(); 976 new UnreachableInst(BB->getContext(), Pred); 977 978 // The loop is now broken, remove it from LI. 979 RemoveLoopFromHierarchy(L); 980 981 // Reprocess the header, which now IS dead. 982 RemoveBlockIfDead(BB, Worklist, L); 983 return; 984 } 985 986 // If pred ends in a uncond branch, add uncond branch to worklist so that 987 // the two blocks will get merged. 988 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 989 if (BI->isUnconditional()) 990 Worklist.push_back(BI); 991 } 992 return; 993 } 994 995 DEBUG(dbgs() << "Nuking dead block: " << *BB); 996 997 // Remove the instructions in the basic block from the worklist. 998 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 999 RemoveFromWorklist(I, Worklist); 1000 1001 // Anything that uses the instructions in this basic block should have their 1002 // uses replaced with undefs. 1003 // If I is not void type then replaceAllUsesWith undef. 1004 // This allows ValueHandlers and custom metadata to adjust itself. 1005 if (!I->getType()->isVoidTy()) 1006 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1007 } 1008 1009 // If this is the edge to the header block for a loop, remove the loop and 1010 // promote all subloops. 1011 if (Loop *BBLoop = LI->getLoopFor(BB)) { 1012 if (BBLoop->getLoopLatch() == BB) { 1013 RemoveLoopFromHierarchy(BBLoop); 1014 if (currentLoop == BBLoop) { 1015 currentLoop = 0; 1016 redoLoop = false; 1017 } 1018 } 1019 } 1020 1021 // Remove the block from the loop info, which removes it from any loops it 1022 // was in. 1023 LI->removeBlock(BB); 1024 1025 1026 // Remove phi node entries in successors for this block. 1027 TerminatorInst *TI = BB->getTerminator(); 1028 SmallVector<BasicBlock*, 4> Succs; 1029 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1030 Succs.push_back(TI->getSuccessor(i)); 1031 TI->getSuccessor(i)->removePredecessor(BB); 1032 } 1033 1034 // Unique the successors, remove anything with multiple uses. 1035 array_pod_sort(Succs.begin(), Succs.end()); 1036 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 1037 1038 // Remove the basic block, including all of the instructions contained in it. 1039 LPM->deleteSimpleAnalysisValue(BB, L); 1040 BB->eraseFromParent(); 1041 // Remove successor blocks here that are not dead, so that we know we only 1042 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 1043 // then getting removed before we revisit them, which is badness. 1044 // 1045 for (unsigned i = 0; i != Succs.size(); ++i) 1046 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 1047 // One exception is loop headers. If this block was the preheader for a 1048 // loop, then we DO want to visit the loop so the loop gets deleted. 1049 // We know that if the successor is a loop header, that this loop had to 1050 // be the preheader: the case where this was the latch block was handled 1051 // above and headers can only have two predecessors. 1052 if (!LI->isLoopHeader(Succs[i])) { 1053 Succs.erase(Succs.begin()+i); 1054 --i; 1055 } 1056 } 1057 1058 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 1059 RemoveBlockIfDead(Succs[i], Worklist, L); 1060 } 1061 1062 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has 1063 /// become unwrapped, either because the backedge was deleted, or because the 1064 /// edge into the header was removed. If the edge into the header from the 1065 /// latch block was removed, the loop is unwrapped but subloops are still alive, 1066 /// so they just reparent loops. If the loops are actually dead, they will be 1067 /// removed later. 1068 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 1069 LPM->deleteLoopFromQueue(L); 1070 RemoveLoopFromWorklist(L); 1071 } 1072 1073 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 1074 // the value specified by Val in the specified loop, or we know it does NOT have 1075 // that value. Rewrite any uses of LIC or of properties correlated to it. 1076 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 1077 Constant *Val, 1078 bool IsEqual) { 1079 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 1080 1081 // FIXME: Support correlated properties, like: 1082 // for (...) 1083 // if (li1 < li2) 1084 // ... 1085 // if (li1 > li2) 1086 // ... 1087 1088 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 1089 // selects, switches. 1090 std::vector<Instruction*> Worklist; 1091 LLVMContext &Context = Val->getContext(); 1092 1093 1094 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 1095 // in the loop with the appropriate one directly. 1096 if (IsEqual || (isa<ConstantInt>(Val) && 1097 Val->getType()->isIntegerTy(1))) { 1098 Value *Replacement; 1099 if (IsEqual) 1100 Replacement = Val; 1101 else 1102 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 1103 !cast<ConstantInt>(Val)->getZExtValue()); 1104 1105 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); 1106 UI != E; ++UI) { 1107 Instruction *U = dyn_cast<Instruction>(*UI); 1108 if (!U || !L->contains(U)) 1109 continue; 1110 Worklist.push_back(U); 1111 } 1112 1113 for (std::vector<Instruction*>::iterator UI = Worklist.begin(); 1114 UI != Worklist.end(); ++UI) 1115 (*UI)->replaceUsesOfWith(LIC, Replacement); 1116 1117 SimplifyCode(Worklist, L); 1118 return; 1119 } 1120 1121 // Otherwise, we don't know the precise value of LIC, but we do know that it 1122 // is certainly NOT "Val". As such, simplify any uses in the loop that we 1123 // can. This case occurs when we unswitch switch statements. 1124 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); 1125 UI != E; ++UI) { 1126 Instruction *U = dyn_cast<Instruction>(*UI); 1127 if (!U || !L->contains(U)) 1128 continue; 1129 1130 Worklist.push_back(U); 1131 1132 // TODO: We could do other simplifications, for example, turning 1133 // 'icmp eq LIC, Val' -> false. 1134 1135 // If we know that LIC is not Val, use this info to simplify code. 1136 SwitchInst *SI = dyn_cast<SwitchInst>(U); 1137 if (SI == 0 || !isa<ConstantInt>(Val)) continue; 1138 1139 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val)); 1140 // Default case is live for multiple values. 1141 if (DeadCase == SI->case_default()) continue; 1142 1143 // Found a dead case value. Don't remove PHI nodes in the 1144 // successor if they become single-entry, those PHI nodes may 1145 // be in the Users list. 1146 1147 BasicBlock *Switch = SI->getParent(); 1148 BasicBlock *SISucc = DeadCase.getCaseSuccessor(); 1149 BasicBlock *Latch = L->getLoopLatch(); 1150 1151 BranchesInfo.setUnswitched(SI, Val); 1152 1153 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical. 1154 // If the DeadCase successor dominates the loop latch, then the 1155 // transformation isn't safe since it will delete the sole predecessor edge 1156 // to the latch. 1157 if (Latch && DT->dominates(SISucc, Latch)) 1158 continue; 1159 1160 // FIXME: This is a hack. We need to keep the successor around 1161 // and hooked up so as to preserve the loop structure, because 1162 // trying to update it is complicated. So instead we preserve the 1163 // loop structure and put the block on a dead code path. 1164 SplitEdge(Switch, SISucc, this); 1165 // Compute the successors instead of relying on the return value 1166 // of SplitEdge, since it may have split the switch successor 1167 // after PHI nodes. 1168 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor(); 1169 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 1170 // Create an "unreachable" destination. 1171 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 1172 Switch->getParent(), 1173 OldSISucc); 1174 new UnreachableInst(Context, Abort); 1175 // Force the new case destination to branch to the "unreachable" 1176 // block while maintaining a (dead) CFG edge to the old block. 1177 NewSISucc->getTerminator()->eraseFromParent(); 1178 BranchInst::Create(Abort, OldSISucc, 1179 ConstantInt::getTrue(Context), NewSISucc); 1180 // Release the PHI operands for this edge. 1181 for (BasicBlock::iterator II = NewSISucc->begin(); 1182 PHINode *PN = dyn_cast<PHINode>(II); ++II) 1183 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 1184 UndefValue::get(PN->getType())); 1185 // Tell the domtree about the new block. We don't fully update the 1186 // domtree here -- instead we force it to do a full recomputation 1187 // after the pass is complete -- but we do need to inform it of 1188 // new blocks. 1189 if (DT) 1190 DT->addNewBlock(Abort, NewSISucc); 1191 } 1192 1193 SimplifyCode(Worklist, L); 1194 } 1195 1196 /// SimplifyCode - Okay, now that we have simplified some instructions in the 1197 /// loop, walk over it and constant prop, dce, and fold control flow where 1198 /// possible. Note that this is effectively a very simple loop-structure-aware 1199 /// optimizer. During processing of this loop, L could very well be deleted, so 1200 /// it must not be used. 1201 /// 1202 /// FIXME: When the loop optimizer is more mature, separate this out to a new 1203 /// pass. 1204 /// 1205 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 1206 while (!Worklist.empty()) { 1207 Instruction *I = Worklist.back(); 1208 Worklist.pop_back(); 1209 1210 // Simple DCE. 1211 if (isInstructionTriviallyDead(I)) { 1212 DEBUG(dbgs() << "Remove dead instruction '" << *I); 1213 1214 // Add uses to the worklist, which may be dead now. 1215 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1216 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1217 Worklist.push_back(Use); 1218 LPM->deleteSimpleAnalysisValue(I, L); 1219 RemoveFromWorklist(I, Worklist); 1220 I->eraseFromParent(); 1221 ++NumSimplify; 1222 continue; 1223 } 1224 1225 // See if instruction simplification can hack this up. This is common for 1226 // things like "select false, X, Y" after unswitching made the condition be 1227 // 'false'. 1228 if (Value *V = SimplifyInstruction(I, 0, 0, DT)) 1229 if (LI->replacementPreservesLCSSAForm(I, V)) { 1230 ReplaceUsesOfWith(I, V, Worklist, L, LPM); 1231 continue; 1232 } 1233 1234 // Special case hacks that appear commonly in unswitched code. 1235 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1236 if (BI->isUnconditional()) { 1237 // If BI's parent is the only pred of the successor, fold the two blocks 1238 // together. 1239 BasicBlock *Pred = BI->getParent(); 1240 BasicBlock *Succ = BI->getSuccessor(0); 1241 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1242 if (!SinglePred) continue; // Nothing to do. 1243 assert(SinglePred == Pred && "CFG broken"); 1244 1245 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1246 << Succ->getName() << "\n"); 1247 1248 // Resolve any single entry PHI nodes in Succ. 1249 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1250 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1251 1252 // If Succ has any successors with PHI nodes, update them to have 1253 // entries coming from Pred instead of Succ. 1254 Succ->replaceAllUsesWith(Pred); 1255 1256 // Move all of the successor contents from Succ to Pred. 1257 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1258 Succ->end()); 1259 LPM->deleteSimpleAnalysisValue(BI, L); 1260 BI->eraseFromParent(); 1261 RemoveFromWorklist(BI, Worklist); 1262 1263 // Remove Succ from the loop tree. 1264 LI->removeBlock(Succ); 1265 LPM->deleteSimpleAnalysisValue(Succ, L); 1266 Succ->eraseFromParent(); 1267 ++NumSimplify; 1268 continue; 1269 } 1270 1271 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1272 // Conditional branch. Turn it into an unconditional branch, then 1273 // remove dead blocks. 1274 continue; // FIXME: Enable. 1275 1276 DEBUG(dbgs() << "Folded branch: " << *BI); 1277 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1278 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1279 DeadSucc->removePredecessor(BI->getParent(), true); 1280 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1281 LPM->deleteSimpleAnalysisValue(BI, L); 1282 BI->eraseFromParent(); 1283 RemoveFromWorklist(BI, Worklist); 1284 ++NumSimplify; 1285 1286 RemoveBlockIfDead(DeadSucc, Worklist, L); 1287 } 1288 continue; 1289 } 1290 } 1291 } 1292