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 #include "llvm/Transforms/Scalar.h" 30 #include "llvm/ADT/STLExtras.h" 31 #include "llvm/ADT/SmallPtrSet.h" 32 #include "llvm/ADT/Statistic.h" 33 #include "llvm/Analysis/GlobalsModRef.h" 34 #include "llvm/Analysis/AssumptionCache.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/ScalarEvolution.h" 40 #include "llvm/Analysis/TargetTransformInfo.h" 41 #include "llvm/Analysis/BlockFrequencyInfoImpl.h" 42 #include "llvm/Analysis/BlockFrequencyInfo.h" 43 #include "llvm/Analysis/BranchProbabilityInfo.h" 44 #include "llvm/Support/BranchProbability.h" 45 #include "llvm/IR/Constants.h" 46 #include "llvm/IR/DerivedTypes.h" 47 #include "llvm/IR/Dominators.h" 48 #include "llvm/IR/Function.h" 49 #include "llvm/IR/Instructions.h" 50 #include "llvm/IR/Module.h" 51 #include "llvm/IR/MDBuilder.h" 52 #include "llvm/Support/CommandLine.h" 53 #include "llvm/Support/Debug.h" 54 #include "llvm/Support/raw_ostream.h" 55 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 56 #include "llvm/Transforms/Utils/Cloning.h" 57 #include "llvm/Transforms/Utils/Local.h" 58 #include "llvm/Transforms/Utils/LoopUtils.h" 59 #include <algorithm> 60 #include <map> 61 #include <set> 62 using namespace llvm; 63 64 #define DEBUG_TYPE "loop-unswitch" 65 66 STATISTIC(NumBranches, "Number of branches unswitched"); 67 STATISTIC(NumSwitches, "Number of switches unswitched"); 68 STATISTIC(NumGuards, "Number of guards unswitched"); 69 STATISTIC(NumSelects , "Number of selects unswitched"); 70 STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 71 STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 72 STATISTIC(TotalInsts, "Total number of instructions analyzed"); 73 74 // The specific value of 100 here was chosen based only on intuition and a 75 // few specific examples. 76 static cl::opt<unsigned> 77 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 78 cl::init(100), cl::Hidden); 79 80 static cl::opt<bool> 81 LoopUnswitchWithBlockFrequency("loop-unswitch-with-block-frequency", 82 cl::init(false), cl::Hidden, 83 cl::desc("Enable the use of the block frequency analysis to access PGO " 84 "heuristics to minimize code growth in cold regions.")); 85 86 static cl::opt<unsigned> 87 ColdnessThreshold("loop-unswitch-coldness-threshold", cl::init(1), cl::Hidden, 88 cl::desc("Coldness threshold in percentage. The loop header frequency " 89 "(relative to the entry frequency) is compared with this " 90 "threshold to determine if non-trivial unswitching should be " 91 "enabled.")); 92 93 namespace { 94 95 class LUAnalysisCache { 96 97 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> > 98 UnswitchedValsMap; 99 100 typedef UnswitchedValsMap::iterator UnswitchedValsIt; 101 102 struct LoopProperties { 103 unsigned CanBeUnswitchedCount; 104 unsigned WasUnswitchedCount; 105 unsigned SizeEstimation; 106 UnswitchedValsMap UnswitchedVals; 107 }; 108 109 // Here we use std::map instead of DenseMap, since we need to keep valid 110 // LoopProperties pointer for current loop for better performance. 111 typedef std::map<const Loop*, LoopProperties> LoopPropsMap; 112 typedef LoopPropsMap::iterator LoopPropsMapIt; 113 114 LoopPropsMap LoopsProperties; 115 UnswitchedValsMap *CurLoopInstructions; 116 LoopProperties *CurrentLoopProperties; 117 118 // A loop unswitching with an estimated cost above this threshold 119 // is not performed. MaxSize is turned into unswitching quota for 120 // the current loop, and reduced correspondingly, though note that 121 // the quota is returned by releaseMemory() when the loop has been 122 // processed, so that MaxSize will return to its previous 123 // value. So in most cases MaxSize will equal the Threshold flag 124 // when a new loop is processed. An exception to that is that 125 // MaxSize will have a smaller value while processing nested loops 126 // that were introduced due to loop unswitching of an outer loop. 127 // 128 // FIXME: The way that MaxSize works is subtle and depends on the 129 // pass manager processing loops and calling releaseMemory() in a 130 // specific order. It would be good to find a more straightforward 131 // way of doing what MaxSize does. 132 unsigned MaxSize; 133 134 public: 135 LUAnalysisCache() 136 : CurLoopInstructions(nullptr), CurrentLoopProperties(nullptr), 137 MaxSize(Threshold) {} 138 139 // Analyze loop. Check its size, calculate is it possible to unswitch 140 // it. Returns true if we can unswitch this loop. 141 bool countLoop(const Loop *L, const TargetTransformInfo &TTI, 142 AssumptionCache *AC); 143 144 // Clean all data related to given loop. 145 void forgetLoop(const Loop *L); 146 147 // Mark case value as unswitched. 148 // Since SI instruction can be partly unswitched, in order to avoid 149 // extra unswitching in cloned loops keep track all unswitched values. 150 void setUnswitched(const SwitchInst *SI, const Value *V); 151 152 // Check was this case value unswitched before or not. 153 bool isUnswitched(const SwitchInst *SI, const Value *V); 154 155 // Returns true if another unswitching could be done within the cost 156 // threshold. 157 bool CostAllowsUnswitching(); 158 159 // Clone all loop-unswitch related loop properties. 160 // Redistribute unswitching quotas. 161 // Note, that new loop data is stored inside the VMap. 162 void cloneData(const Loop *NewLoop, const Loop *OldLoop, 163 const ValueToValueMapTy &VMap); 164 }; 165 166 class LoopUnswitch : public LoopPass { 167 LoopInfo *LI; // Loop information 168 LPPassManager *LPM; 169 AssumptionCache *AC; 170 171 // Used to check if second loop needs processing after 172 // RewriteLoopBodyWithConditionConstant rewrites first loop. 173 std::vector<Loop*> LoopProcessWorklist; 174 175 LUAnalysisCache BranchesInfo; 176 177 bool EnabledPGO; 178 179 // BFI and ColdEntryFreq are only used when PGO and 180 // LoopUnswitchWithBlockFrequency are enabled. 181 BlockFrequencyInfo BFI; 182 BlockFrequency ColdEntryFreq; 183 184 bool OptimizeForSize; 185 bool redoLoop; 186 187 Loop *currentLoop; 188 DominatorTree *DT; 189 BasicBlock *loopHeader; 190 BasicBlock *loopPreheader; 191 192 bool SanitizeMemory; 193 LoopSafetyInfo SafetyInfo; 194 195 // LoopBlocks contains all of the basic blocks of the loop, including the 196 // preheader of the loop, the body of the loop, and the exit blocks of the 197 // loop, in that order. 198 std::vector<BasicBlock*> LoopBlocks; 199 // NewBlocks contained cloned copy of basic blocks from LoopBlocks. 200 std::vector<BasicBlock*> NewBlocks; 201 202 public: 203 static char ID; // Pass ID, replacement for typeid 204 explicit LoopUnswitch(bool Os = false) : 205 LoopPass(ID), OptimizeForSize(Os), redoLoop(false), 206 currentLoop(nullptr), DT(nullptr), loopHeader(nullptr), 207 loopPreheader(nullptr) { 208 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry()); 209 } 210 211 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 212 bool processCurrentLoop(); 213 214 /// This transformation requires natural loop information & requires that 215 /// loop preheaders be inserted into the CFG. 216 /// 217 void getAnalysisUsage(AnalysisUsage &AU) const override { 218 AU.addRequired<AssumptionCacheTracker>(); 219 AU.addRequired<TargetTransformInfoWrapperPass>(); 220 getLoopAnalysisUsage(AU); 221 } 222 223 private: 224 225 void releaseMemory() override { 226 BranchesInfo.forgetLoop(currentLoop); 227 } 228 229 void initLoopData() { 230 loopHeader = currentLoop->getHeader(); 231 loopPreheader = currentLoop->getLoopPreheader(); 232 } 233 234 /// Split all of the edges from inside the loop to their exit blocks. 235 /// Update the appropriate Phi nodes as we do so. 236 void SplitExitEdges(Loop *L, 237 const SmallVectorImpl<BasicBlock *> &ExitBlocks); 238 239 bool TryTrivialLoopUnswitch(bool &Changed); 240 241 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val, 242 TerminatorInst *TI = nullptr); 243 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 244 BasicBlock *ExitBlock, TerminatorInst *TI); 245 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L, 246 TerminatorInst *TI); 247 248 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 249 Constant *Val, bool isEqual); 250 251 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 252 BasicBlock *TrueDest, 253 BasicBlock *FalseDest, 254 Instruction *InsertPt, 255 TerminatorInst *TI); 256 257 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 258 }; 259 } 260 261 // Analyze loop. Check its size, calculate is it possible to unswitch 262 // it. Returns true if we can unswitch this loop. 263 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI, 264 AssumptionCache *AC) { 265 266 LoopPropsMapIt PropsIt; 267 bool Inserted; 268 std::tie(PropsIt, Inserted) = 269 LoopsProperties.insert(std::make_pair(L, LoopProperties())); 270 271 LoopProperties &Props = PropsIt->second; 272 273 if (Inserted) { 274 // New loop. 275 276 // Limit the number of instructions to avoid causing significant code 277 // expansion, and the number of basic blocks, to avoid loops with 278 // large numbers of branches which cause loop unswitching to go crazy. 279 // This is a very ad-hoc heuristic. 280 281 SmallPtrSet<const Value *, 32> EphValues; 282 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 283 284 // FIXME: This is overly conservative because it does not take into 285 // consideration code simplification opportunities and code that can 286 // be shared by the resultant unswitched loops. 287 CodeMetrics Metrics; 288 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E; 289 ++I) 290 Metrics.analyzeBasicBlock(*I, TTI, EphValues); 291 292 Props.SizeEstimation = Metrics.NumInsts; 293 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation); 294 Props.WasUnswitchedCount = 0; 295 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount; 296 297 if (Metrics.notDuplicatable) { 298 DEBUG(dbgs() << "NOT unswitching loop %" 299 << L->getHeader()->getName() << ", contents cannot be " 300 << "duplicated!\n"); 301 return false; 302 } 303 } 304 305 // Be careful. This links are good only before new loop addition. 306 CurrentLoopProperties = &Props; 307 CurLoopInstructions = &Props.UnswitchedVals; 308 309 return true; 310 } 311 312 // Clean all data related to given loop. 313 void LUAnalysisCache::forgetLoop(const Loop *L) { 314 315 LoopPropsMapIt LIt = LoopsProperties.find(L); 316 317 if (LIt != LoopsProperties.end()) { 318 LoopProperties &Props = LIt->second; 319 MaxSize += (Props.CanBeUnswitchedCount + Props.WasUnswitchedCount) * 320 Props.SizeEstimation; 321 LoopsProperties.erase(LIt); 322 } 323 324 CurrentLoopProperties = nullptr; 325 CurLoopInstructions = nullptr; 326 } 327 328 // Mark case value as unswitched. 329 // Since SI instruction can be partly unswitched, in order to avoid 330 // extra unswitching in cloned loops keep track all unswitched values. 331 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) { 332 (*CurLoopInstructions)[SI].insert(V); 333 } 334 335 // Check was this case value unswitched before or not. 336 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) { 337 return (*CurLoopInstructions)[SI].count(V); 338 } 339 340 bool LUAnalysisCache::CostAllowsUnswitching() { 341 return CurrentLoopProperties->CanBeUnswitchedCount > 0; 342 } 343 344 // Clone all loop-unswitch related loop properties. 345 // Redistribute unswitching quotas. 346 // Note, that new loop data is stored inside the VMap. 347 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop, 348 const ValueToValueMapTy &VMap) { 349 350 LoopProperties &NewLoopProps = LoopsProperties[NewLoop]; 351 LoopProperties &OldLoopProps = *CurrentLoopProperties; 352 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals; 353 354 // Reallocate "can-be-unswitched quota" 355 356 --OldLoopProps.CanBeUnswitchedCount; 357 ++OldLoopProps.WasUnswitchedCount; 358 NewLoopProps.WasUnswitchedCount = 0; 359 unsigned Quota = OldLoopProps.CanBeUnswitchedCount; 360 NewLoopProps.CanBeUnswitchedCount = Quota / 2; 361 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2; 362 363 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation; 364 365 // Clone unswitched values info: 366 // for new loop switches we clone info about values that was 367 // already unswitched and has redundant successors. 368 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) { 369 const SwitchInst *OldInst = I->first; 370 Value *NewI = VMap.lookup(OldInst); 371 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI); 372 assert(NewInst && "All instructions that are in SrcBB must be in VMap."); 373 374 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst]; 375 } 376 } 377 378 char LoopUnswitch::ID = 0; 379 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", 380 false, false) 381 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 382 INITIALIZE_PASS_DEPENDENCY(LoopPass) 383 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 384 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", 385 false, false) 386 387 Pass *llvm::createLoopUnswitchPass(bool Os) { 388 return new LoopUnswitch(Os); 389 } 390 391 /// Cond is a condition that occurs in L. If it is invariant in the loop, or has 392 /// an invariant piece, return the invariant. Otherwise, return null. 393 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed, 394 DenseMap<Value *, Value *> &Cache) { 395 auto CacheIt = Cache.find(Cond); 396 if (CacheIt != Cache.end()) 397 return CacheIt->second; 398 399 // We started analyze new instruction, increment scanned instructions counter. 400 ++TotalInsts; 401 402 // We can never unswitch on vector conditions. 403 if (Cond->getType()->isVectorTy()) 404 return nullptr; 405 406 // Constants should be folded, not unswitched on! 407 if (isa<Constant>(Cond)) return nullptr; 408 409 // TODO: Handle: br (VARIANT|INVARIANT). 410 411 // Hoist simple values out. 412 if (L->makeLoopInvariant(Cond, Changed)) { 413 Cache[Cond] = Cond; 414 return Cond; 415 } 416 417 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 418 if (BO->getOpcode() == Instruction::And || 419 BO->getOpcode() == Instruction::Or) { 420 // If either the left or right side is invariant, we can unswitch on this, 421 // which will cause the branch to go away in one loop and the condition to 422 // simplify in the other one. 423 if (Value *LHS = 424 FindLIVLoopCondition(BO->getOperand(0), L, Changed, Cache)) { 425 Cache[Cond] = LHS; 426 return LHS; 427 } 428 if (Value *RHS = 429 FindLIVLoopCondition(BO->getOperand(1), L, Changed, Cache)) { 430 Cache[Cond] = RHS; 431 return RHS; 432 } 433 } 434 435 Cache[Cond] = nullptr; 436 return nullptr; 437 } 438 439 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 440 DenseMap<Value *, Value *> Cache; 441 return FindLIVLoopCondition(Cond, L, Changed, Cache); 442 } 443 444 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 445 if (skipLoop(L)) 446 return false; 447 448 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache( 449 *L->getHeader()->getParent()); 450 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 451 LPM = &LPM_Ref; 452 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 453 currentLoop = L; 454 Function *F = currentLoop->getHeader()->getParent(); 455 456 SanitizeMemory = F->hasFnAttribute(Attribute::SanitizeMemory); 457 if (SanitizeMemory) 458 computeLoopSafetyInfo(&SafetyInfo, L); 459 460 EnabledPGO = F->getEntryCount().hasValue(); 461 462 if (LoopUnswitchWithBlockFrequency && EnabledPGO) { 463 BranchProbabilityInfo BPI(*F, *LI); 464 BFI.calculate(*L->getHeader()->getParent(), BPI, *LI); 465 466 // Use BranchProbability to compute a minimum frequency based on 467 // function entry baseline frequency. Loops with headers below this 468 // frequency are considered as cold. 469 const BranchProbability ColdProb(ColdnessThreshold, 100); 470 ColdEntryFreq = BlockFrequency(BFI.getEntryFreq()) * ColdProb; 471 } 472 473 bool Changed = false; 474 do { 475 assert(currentLoop->isLCSSAForm(*DT)); 476 redoLoop = false; 477 Changed |= processCurrentLoop(); 478 } while(redoLoop); 479 480 // FIXME: Reconstruct dom info, because it is not preserved properly. 481 if (Changed) 482 DT->recalculate(*F); 483 return Changed; 484 } 485 486 /// Do actual work and unswitch loop if possible and profitable. 487 bool LoopUnswitch::processCurrentLoop() { 488 bool Changed = false; 489 490 initLoopData(); 491 492 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 493 if (!loopPreheader) 494 return false; 495 496 // Loops with indirectbr cannot be cloned. 497 if (!currentLoop->isSafeToClone()) 498 return false; 499 500 // Without dedicated exits, splitting the exit edge may fail. 501 if (!currentLoop->hasDedicatedExits()) 502 return false; 503 504 LLVMContext &Context = loopHeader->getContext(); 505 506 // Analyze loop cost, and stop unswitching if loop content can not be duplicated. 507 if (!BranchesInfo.countLoop( 508 currentLoop, getAnalysis<TargetTransformInfoWrapperPass>().getTTI( 509 *currentLoop->getHeader()->getParent()), 510 AC)) 511 return false; 512 513 // Try trivial unswitch first before loop over other basic blocks in the loop. 514 if (TryTrivialLoopUnswitch(Changed)) { 515 return true; 516 } 517 518 // Run through the instructions in the loop, keeping track of three things: 519 // 520 // - That we do not unswitch loops containing convergent operations, as we 521 // might be making them control dependent on the unswitch value when they 522 // were not before. 523 // FIXME: This could be refined to only bail if the convergent operation is 524 // not already control-dependent on the unswitch value. 525 // 526 // - That basic blocks in the loop contain invokes whose predecessor edges we 527 // cannot split. 528 // 529 // - The set of guard intrinsics encountered (these are non terminator 530 // instructions that are also profitable to be unswitched). 531 532 SmallVector<IntrinsicInst *, 4> Guards; 533 534 for (const auto BB : currentLoop->blocks()) { 535 for (auto &I : *BB) { 536 auto CS = CallSite(&I); 537 if (!CS) continue; 538 if (CS.hasFnAttr(Attribute::Convergent)) 539 return false; 540 if (auto *II = dyn_cast<InvokeInst>(&I)) 541 if (!II->getUnwindDest()->canSplitPredecessors()) 542 return false; 543 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 544 if (II->getIntrinsicID() == Intrinsic::experimental_guard) 545 Guards.push_back(II); 546 } 547 } 548 549 // Do not do non-trivial unswitch while optimizing for size. 550 // FIXME: Use Function::optForSize(). 551 if (OptimizeForSize || 552 loopHeader->getParent()->hasFnAttribute(Attribute::OptimizeForSize)) 553 return false; 554 555 if (LoopUnswitchWithBlockFrequency && EnabledPGO) { 556 // Compute the weighted frequency of the hottest block in the 557 // loop (loopHeader in this case since inner loops should be 558 // processed before outer loop). If it is less than ColdFrequency, 559 // we should not unswitch. 560 BlockFrequency LoopEntryFreq = BFI.getBlockFreq(loopHeader); 561 if (LoopEntryFreq < ColdEntryFreq) 562 return false; 563 } 564 565 for (IntrinsicInst *Guard : Guards) { 566 Value *LoopCond = 567 FindLIVLoopCondition(Guard->getOperand(0), currentLoop, Changed); 568 if (LoopCond && 569 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context))) { 570 // NB! Unswitching (if successful) could have erased some of the 571 // instructions in Guards leaving dangling pointers there. This is fine 572 // because we're returning now, and won't look at Guards again. 573 ++NumGuards; 574 return true; 575 } 576 } 577 578 // Loop over all of the basic blocks in the loop. If we find an interior 579 // block that is branching on a loop-invariant condition, we can unswitch this 580 // loop. 581 for (Loop::block_iterator I = currentLoop->block_begin(), 582 E = currentLoop->block_end(); I != E; ++I) { 583 TerminatorInst *TI = (*I)->getTerminator(); 584 585 // Unswitching on a potentially uninitialized predicate is not 586 // MSan-friendly. Limit this to the cases when the original predicate is 587 // guaranteed to execute, to avoid creating a use-of-uninitialized-value 588 // in the code that did not have one. 589 // This is a workaround for the discrepancy between LLVM IR and MSan 590 // semantics. See PR28054 for more details. 591 if (SanitizeMemory && 592 !isGuaranteedToExecute(*TI, DT, currentLoop, &SafetyInfo)) 593 continue; 594 595 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 596 // If this isn't branching on an invariant condition, we can't unswitch 597 // it. 598 if (BI->isConditional()) { 599 // See if this, or some part of it, is loop invariant. If so, we can 600 // unswitch on it if we desire. 601 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 602 currentLoop, Changed); 603 if (LoopCond && 604 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context), TI)) { 605 ++NumBranches; 606 return true; 607 } 608 } 609 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 610 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 611 currentLoop, Changed); 612 unsigned NumCases = SI->getNumCases(); 613 if (LoopCond && NumCases) { 614 // Find a value to unswitch on: 615 // FIXME: this should chose the most expensive case! 616 // FIXME: scan for a case with a non-critical edge? 617 Constant *UnswitchVal = nullptr; 618 619 // Do not process same value again and again. 620 // At this point we have some cases already unswitched and 621 // some not yet unswitched. Let's find the first not yet unswitched one. 622 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 623 i != e; ++i) { 624 Constant *UnswitchValCandidate = i.getCaseValue(); 625 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) { 626 UnswitchVal = UnswitchValCandidate; 627 break; 628 } 629 } 630 631 if (!UnswitchVal) 632 continue; 633 634 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 635 ++NumSwitches; 636 return true; 637 } 638 } 639 } 640 641 // Scan the instructions to check for unswitchable values. 642 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 643 BBI != E; ++BBI) 644 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 645 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 646 currentLoop, Changed); 647 if (LoopCond && UnswitchIfProfitable(LoopCond, 648 ConstantInt::getTrue(Context))) { 649 ++NumSelects; 650 return true; 651 } 652 } 653 } 654 return Changed; 655 } 656 657 /// Check to see if all paths from BB exit the loop with no side effects 658 /// (including infinite loops). 659 /// 660 /// If true, we return true and set ExitBB to the block we 661 /// exit through. 662 /// 663 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 664 BasicBlock *&ExitBB, 665 std::set<BasicBlock*> &Visited) { 666 if (!Visited.insert(BB).second) { 667 // Already visited. Without more analysis, this could indicate an infinite 668 // loop. 669 return false; 670 } 671 if (!L->contains(BB)) { 672 // Otherwise, this is a loop exit, this is fine so long as this is the 673 // first exit. 674 if (ExitBB) return false; 675 ExitBB = BB; 676 return true; 677 } 678 679 // Otherwise, this is an unvisited intra-loop node. Check all successors. 680 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 681 // Check to see if the successor is a trivial loop exit. 682 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 683 return false; 684 } 685 686 // Okay, everything after this looks good, check to make sure that this block 687 // doesn't include any side effects. 688 for (Instruction &I : *BB) 689 if (I.mayHaveSideEffects()) 690 return false; 691 692 return true; 693 } 694 695 /// Return true if the specified block unconditionally leads to an exit from 696 /// the specified loop, and has no side-effects in the process. If so, return 697 /// the block that is exited to, otherwise return null. 698 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 699 std::set<BasicBlock*> Visited; 700 Visited.insert(L->getHeader()); // Branches to header make infinite loops. 701 BasicBlock *ExitBB = nullptr; 702 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 703 return ExitBB; 704 return nullptr; 705 } 706 707 /// We have found that we can unswitch currentLoop when LoopCond == Val to 708 /// simplify the loop. If we decide that this is profitable, 709 /// unswitch the loop, reprocess the pieces, then return true. 710 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val, 711 TerminatorInst *TI) { 712 // Check to see if it would be profitable to unswitch current loop. 713 if (!BranchesInfo.CostAllowsUnswitching()) { 714 DEBUG(dbgs() << "NOT unswitching loop %" 715 << currentLoop->getHeader()->getName() 716 << " at non-trivial condition '" << *Val 717 << "' == " << *LoopCond << "\n" 718 << ". Cost too high.\n"); 719 return false; 720 } 721 722 UnswitchNontrivialCondition(LoopCond, Val, currentLoop, TI); 723 return true; 724 } 725 726 /// Recursively clone the specified loop and all of its children, 727 /// mapping the blocks with the specified map. 728 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, 729 LoopInfo *LI, LPPassManager *LPM) { 730 Loop &New = LPM->addLoop(PL); 731 732 // Add all of the blocks in L to the new loop. 733 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 734 I != E; ++I) 735 if (LI->getLoopFor(*I) == L) 736 New.addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI); 737 738 // Add all of the subloops to the new loop. 739 for (Loop *I : *L) 740 CloneLoop(I, &New, VM, LI, LPM); 741 742 return &New; 743 } 744 745 static void copyMetadata(Instruction *DstInst, const Instruction *SrcInst, 746 bool Swapped) { 747 if (!SrcInst || !SrcInst->hasMetadata()) 748 return; 749 750 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs; 751 SrcInst->getAllMetadata(MDs); 752 for (auto &MD : MDs) { 753 switch (MD.first) { 754 default: 755 break; 756 case LLVMContext::MD_prof: 757 if (Swapped && MD.second->getNumOperands() == 3 && 758 isa<MDString>(MD.second->getOperand(0))) { 759 MDString *MDName = cast<MDString>(MD.second->getOperand(0)); 760 if (MDName->getString() == "branch_weights") { 761 auto *ValT = cast_or_null<ConstantAsMetadata>( 762 MD.second->getOperand(1))->getValue(); 763 auto *ValF = cast_or_null<ConstantAsMetadata>( 764 MD.second->getOperand(2))->getValue(); 765 assert(ValT && ValF && "Invalid Operands of branch_weights"); 766 auto NewMD = 767 MDBuilder(DstInst->getParent()->getContext()) 768 .createBranchWeights(cast<ConstantInt>(ValF)->getZExtValue(), 769 cast<ConstantInt>(ValT)->getZExtValue()); 770 MD.second = NewMD; 771 } 772 } 773 // fallthrough. 774 case LLVMContext::MD_make_implicit: 775 case LLVMContext::MD_dbg: 776 DstInst->setMetadata(MD.first, MD.second); 777 } 778 } 779 } 780 781 /// Emit a conditional branch on two values if LIC == Val, branch to TrueDst, 782 /// otherwise branch to FalseDest. Insert the code immediately before InsertPt. 783 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 784 BasicBlock *TrueDest, 785 BasicBlock *FalseDest, 786 Instruction *InsertPt, 787 TerminatorInst *TI) { 788 // Insert a conditional branch on LIC to the two preheaders. The original 789 // code is the true version and the new code is the false version. 790 Value *BranchVal = LIC; 791 bool Swapped = false; 792 if (!isa<ConstantInt>(Val) || 793 Val->getType() != Type::getInt1Ty(LIC->getContext())) 794 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val); 795 else if (Val != ConstantInt::getTrue(Val->getContext())) { 796 // We want to enter the new loop when the condition is true. 797 std::swap(TrueDest, FalseDest); 798 Swapped = true; 799 } 800 801 // Insert the new branch. 802 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 803 copyMetadata(BI, TI, Swapped); 804 805 // If either edge is critical, split it. This helps preserve LoopSimplify 806 // form for enclosing loops. 807 auto Options = CriticalEdgeSplittingOptions(DT, LI).setPreserveLCSSA(); 808 SplitCriticalEdge(BI, 0, Options); 809 SplitCriticalEdge(BI, 1, Options); 810 } 811 812 /// Given a loop that has a trivial unswitchable condition in it (a cond branch 813 /// from its header block to its latch block, where the path through the loop 814 /// that doesn't execute its body has no side-effects), unswitch it. This 815 /// doesn't involve any code duplication, just moving the conditional branch 816 /// outside of the loop and updating loop info. 817 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 818 BasicBlock *ExitBlock, 819 TerminatorInst *TI) { 820 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 821 << loopHeader->getName() << " [" << L->getBlocks().size() 822 << " blocks] in Function " 823 << L->getHeader()->getParent()->getName() << " on cond: " << *Val 824 << " == " << *Cond << "\n"); 825 826 // First step, split the preheader, so that we know that there is a safe place 827 // to insert the conditional branch. We will change loopPreheader to have a 828 // conditional branch on Cond. 829 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, DT, LI); 830 831 // Now that we have a place to insert the conditional branch, create a place 832 // to branch to: this is the exit block out of the loop that we should 833 // short-circuit to. 834 835 // Split this block now, so that the loop maintains its exit block, and so 836 // that the jump from the preheader can execute the contents of the exit block 837 // without actually branching to it (the exit block should be dominated by the 838 // loop header, not the preheader). 839 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 840 BasicBlock *NewExit = SplitBlock(ExitBlock, &ExitBlock->front(), DT, LI); 841 842 // Okay, now we have a position to branch from and a position to branch to, 843 // insert the new conditional branch. 844 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 845 loopPreheader->getTerminator(), TI); 846 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 847 loopPreheader->getTerminator()->eraseFromParent(); 848 849 // We need to reprocess this loop, it could be unswitched again. 850 redoLoop = true; 851 852 // Now that we know that the loop is never entered when this condition is a 853 // particular value, rewrite the loop with this info. We know that this will 854 // at least eliminate the old branch. 855 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 856 ++NumTrivial; 857 } 858 859 /// Check if the first non-constant condition starting from the loop header is 860 /// a trivial unswitch condition: that is, a condition controls whether or not 861 /// the loop does anything at all. If it is a trivial condition, unswitching 862 /// produces no code duplications (equivalently, it produces a simpler loop and 863 /// a new empty loop, which gets deleted). Therefore always unswitch trivial 864 /// condition. 865 bool LoopUnswitch::TryTrivialLoopUnswitch(bool &Changed) { 866 BasicBlock *CurrentBB = currentLoop->getHeader(); 867 TerminatorInst *CurrentTerm = CurrentBB->getTerminator(); 868 LLVMContext &Context = CurrentBB->getContext(); 869 870 // If loop header has only one reachable successor (currently via an 871 // unconditional branch or constant foldable conditional branch, but 872 // should also consider adding constant foldable switch instruction in 873 // future), we should keep looking for trivial condition candidates in 874 // the successor as well. An alternative is to constant fold conditions 875 // and merge successors into loop header (then we only need to check header's 876 // terminator). The reason for not doing this in LoopUnswitch pass is that 877 // it could potentially break LoopPassManager's invariants. Folding dead 878 // branches could either eliminate the current loop or make other loops 879 // unreachable. LCSSA form might also not be preserved after deleting 880 // branches. The following code keeps traversing loop header's successors 881 // until it finds the trivial condition candidate (condition that is not a 882 // constant). Since unswitching generates branches with constant conditions, 883 // this scenario could be very common in practice. 884 SmallSet<BasicBlock*, 8> Visited; 885 886 while (true) { 887 // If we exit loop or reach a previous visited block, then 888 // we can not reach any trivial condition candidates (unfoldable 889 // branch instructions or switch instructions) and no unswitch 890 // can happen. Exit and return false. 891 if (!currentLoop->contains(CurrentBB) || !Visited.insert(CurrentBB).second) 892 return false; 893 894 // Check if this loop will execute any side-effecting instructions (e.g. 895 // stores, calls, volatile loads) in the part of the loop that the code 896 // *would* execute. Check the header first. 897 for (Instruction &I : *CurrentBB) 898 if (I.mayHaveSideEffects()) 899 return false; 900 901 // FIXME: add check for constant foldable switch instructions. 902 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) { 903 if (BI->isUnconditional()) { 904 CurrentBB = BI->getSuccessor(0); 905 } else if (BI->getCondition() == ConstantInt::getTrue(Context)) { 906 CurrentBB = BI->getSuccessor(0); 907 } else if (BI->getCondition() == ConstantInt::getFalse(Context)) { 908 CurrentBB = BI->getSuccessor(1); 909 } else { 910 // Found a trivial condition candidate: non-foldable conditional branch. 911 break; 912 } 913 } else { 914 break; 915 } 916 917 CurrentTerm = CurrentBB->getTerminator(); 918 } 919 920 // CondVal is the condition that controls the trivial condition. 921 // LoopExitBB is the BasicBlock that loop exits when meets trivial condition. 922 Constant *CondVal = nullptr; 923 BasicBlock *LoopExitBB = nullptr; 924 925 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) { 926 // If this isn't branching on an invariant condition, we can't unswitch it. 927 if (!BI->isConditional()) 928 return false; 929 930 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 931 currentLoop, Changed); 932 933 // Unswitch only if the trivial condition itself is an LIV (not 934 // partial LIV which could occur in and/or) 935 if (!LoopCond || LoopCond != BI->getCondition()) 936 return false; 937 938 // Check to see if a successor of the branch is guaranteed to 939 // exit through a unique exit block without having any 940 // side-effects. If so, determine the value of Cond that causes 941 // it to do this. 942 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 943 BI->getSuccessor(0)))) { 944 CondVal = ConstantInt::getTrue(Context); 945 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 946 BI->getSuccessor(1)))) { 947 CondVal = ConstantInt::getFalse(Context); 948 } 949 950 // If we didn't find a single unique LoopExit block, or if the loop exit 951 // block contains phi nodes, this isn't trivial. 952 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 953 return false; // Can't handle this. 954 955 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB, 956 CurrentTerm); 957 ++NumBranches; 958 return true; 959 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurrentTerm)) { 960 // If this isn't switching on an invariant condition, we can't unswitch it. 961 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 962 currentLoop, Changed); 963 964 // Unswitch only if the trivial condition itself is an LIV (not 965 // partial LIV which could occur in and/or) 966 if (!LoopCond || LoopCond != SI->getCondition()) 967 return false; 968 969 // Check to see if a successor of the switch is guaranteed to go to the 970 // latch block or exit through a one exit block without having any 971 // side-effects. If so, determine the value of Cond that causes it to do 972 // this. 973 // Note that we can't trivially unswitch on the default case or 974 // on already unswitched cases. 975 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 976 i != e; ++i) { 977 BasicBlock *LoopExitCandidate; 978 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop, 979 i.getCaseSuccessor()))) { 980 // Okay, we found a trivial case, remember the value that is trivial. 981 ConstantInt *CaseVal = i.getCaseValue(); 982 983 // Check that it was not unswitched before, since already unswitched 984 // trivial vals are looks trivial too. 985 if (BranchesInfo.isUnswitched(SI, CaseVal)) 986 continue; 987 LoopExitBB = LoopExitCandidate; 988 CondVal = CaseVal; 989 break; 990 } 991 } 992 993 // If we didn't find a single unique LoopExit block, or if the loop exit 994 // block contains phi nodes, this isn't trivial. 995 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 996 return false; // Can't handle this. 997 998 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB, 999 nullptr); 1000 ++NumSwitches; 1001 return true; 1002 } 1003 return false; 1004 } 1005 1006 /// Split all of the edges from inside the loop to their exit blocks. 1007 /// Update the appropriate Phi nodes as we do so. 1008 void LoopUnswitch::SplitExitEdges(Loop *L, 1009 const SmallVectorImpl<BasicBlock *> &ExitBlocks){ 1010 1011 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 1012 BasicBlock *ExitBlock = ExitBlocks[i]; 1013 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 1014 pred_end(ExitBlock)); 1015 1016 // Although SplitBlockPredecessors doesn't preserve loop-simplify in 1017 // general, if we call it on all predecessors of all exits then it does. 1018 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", DT, LI, 1019 /*PreserveLCSSA*/ true); 1020 } 1021 } 1022 1023 /// We determined that the loop is profitable to unswitch when LIC equal Val. 1024 /// Split it into loop versions and test the condition outside of either loop. 1025 /// Return the loops created as Out1/Out2. 1026 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 1027 Loop *L, TerminatorInst *TI) { 1028 Function *F = loopHeader->getParent(); 1029 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 1030 << loopHeader->getName() << " [" << L->getBlocks().size() 1031 << " blocks] in Function " << F->getName() 1032 << " when '" << *Val << "' == " << *LIC << "\n"); 1033 1034 if (auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>()) 1035 SEWP->getSE().forgetLoop(L); 1036 1037 LoopBlocks.clear(); 1038 NewBlocks.clear(); 1039 1040 // First step, split the preheader and exit blocks, and add these blocks to 1041 // the LoopBlocks list. 1042 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, DT, LI); 1043 LoopBlocks.push_back(NewPreheader); 1044 1045 // We want the loop to come after the preheader, but before the exit blocks. 1046 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 1047 1048 SmallVector<BasicBlock*, 8> ExitBlocks; 1049 L->getUniqueExitBlocks(ExitBlocks); 1050 1051 // Split all of the edges from inside the loop to their exit blocks. Update 1052 // the appropriate Phi nodes as we do so. 1053 SplitExitEdges(L, ExitBlocks); 1054 1055 // The exit blocks may have been changed due to edge splitting, recompute. 1056 ExitBlocks.clear(); 1057 L->getUniqueExitBlocks(ExitBlocks); 1058 1059 // Add exit blocks to the loop blocks. 1060 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 1061 1062 // Next step, clone all of the basic blocks that make up the loop (including 1063 // the loop preheader and exit blocks), keeping track of the mapping between 1064 // the instructions and blocks. 1065 NewBlocks.reserve(LoopBlocks.size()); 1066 ValueToValueMapTy VMap; 1067 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 1068 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); 1069 1070 NewBlocks.push_back(NewBB); 1071 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. 1072 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); 1073 } 1074 1075 // Splice the newly inserted blocks into the function right before the 1076 // original preheader. 1077 F->getBasicBlockList().splice(NewPreheader->getIterator(), 1078 F->getBasicBlockList(), 1079 NewBlocks[0]->getIterator(), F->end()); 1080 1081 // FIXME: We could register any cloned assumptions instead of clearing the 1082 // whole function's cache. 1083 AC->clear(); 1084 1085 // Now we create the new Loop object for the versioned loop. 1086 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); 1087 1088 // Recalculate unswitching quota, inherit simplified switches info for NewBB, 1089 // Probably clone more loop-unswitch related loop properties. 1090 BranchesInfo.cloneData(NewLoop, L, VMap); 1091 1092 Loop *ParentLoop = L->getParentLoop(); 1093 if (ParentLoop) { 1094 // Make sure to add the cloned preheader and exit blocks to the parent loop 1095 // as well. 1096 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); 1097 } 1098 1099 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 1100 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]); 1101 // The new exit block should be in the same loop as the old one. 1102 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 1103 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); 1104 1105 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 1106 "Exit block should have been split to have one successor!"); 1107 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 1108 1109 // If the successor of the exit block had PHI nodes, add an entry for 1110 // NewExit. 1111 for (BasicBlock::iterator I = ExitSucc->begin(); 1112 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1113 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 1114 ValueToValueMapTy::iterator It = VMap.find(V); 1115 if (It != VMap.end()) V = It->second; 1116 PN->addIncoming(V, NewExit); 1117 } 1118 1119 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) { 1120 PHINode *PN = PHINode::Create(LPad->getType(), 0, "", 1121 &*ExitSucc->getFirstInsertionPt()); 1122 1123 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc); 1124 I != E; ++I) { 1125 BasicBlock *BB = *I; 1126 LandingPadInst *LPI = BB->getLandingPadInst(); 1127 LPI->replaceAllUsesWith(PN); 1128 PN->addIncoming(LPI, BB); 1129 } 1130 } 1131 } 1132 1133 // Rewrite the code to refer to itself. 1134 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 1135 for (Instruction &I : *NewBlocks[i]) 1136 RemapInstruction(&I, VMap, 1137 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1138 1139 // Rewrite the original preheader to select between versions of the loop. 1140 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 1141 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 1142 "Preheader splitting did not work correctly!"); 1143 1144 // Emit the new branch that selects between the two versions of this loop. 1145 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR, 1146 TI); 1147 LPM->deleteSimpleAnalysisValue(OldBR, L); 1148 OldBR->eraseFromParent(); 1149 1150 LoopProcessWorklist.push_back(NewLoop); 1151 redoLoop = true; 1152 1153 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody 1154 // deletes the instruction (for example by simplifying a PHI that feeds into 1155 // the condition that we're unswitching on), we don't rewrite the second 1156 // iteration. 1157 WeakVH LICHandle(LIC); 1158 1159 // Now we rewrite the original code to know that the condition is true and the 1160 // new code to know that the condition is false. 1161 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); 1162 1163 // It's possible that simplifying one loop could cause the other to be 1164 // changed to another value or a constant. If its a constant, don't simplify 1165 // it. 1166 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop && 1167 LICHandle && !isa<Constant>(LICHandle)) 1168 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); 1169 } 1170 1171 /// Remove all instances of I from the worklist vector specified. 1172 static void RemoveFromWorklist(Instruction *I, 1173 std::vector<Instruction*> &Worklist) { 1174 1175 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I), 1176 Worklist.end()); 1177 } 1178 1179 /// When we find that I really equals V, remove I from the 1180 /// program, replacing all uses with V and update the worklist. 1181 static void ReplaceUsesOfWith(Instruction *I, Value *V, 1182 std::vector<Instruction*> &Worklist, 1183 Loop *L, LPPassManager *LPM) { 1184 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 1185 1186 // Add uses to the worklist, which may be dead now. 1187 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1188 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1189 Worklist.push_back(Use); 1190 1191 // Add users to the worklist which may be simplified now. 1192 for (User *U : I->users()) 1193 Worklist.push_back(cast<Instruction>(U)); 1194 LPM->deleteSimpleAnalysisValue(I, L); 1195 RemoveFromWorklist(I, Worklist); 1196 I->replaceAllUsesWith(V); 1197 I->eraseFromParent(); 1198 ++NumSimplify; 1199 } 1200 1201 /// We know either that the value LIC has the value specified by Val in the 1202 /// specified loop, or we know it does NOT have that value. 1203 /// Rewrite any uses of LIC or of properties correlated to it. 1204 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 1205 Constant *Val, 1206 bool IsEqual) { 1207 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 1208 1209 // FIXME: Support correlated properties, like: 1210 // for (...) 1211 // if (li1 < li2) 1212 // ... 1213 // if (li1 > li2) 1214 // ... 1215 1216 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 1217 // selects, switches. 1218 std::vector<Instruction*> Worklist; 1219 LLVMContext &Context = Val->getContext(); 1220 1221 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 1222 // in the loop with the appropriate one directly. 1223 if (IsEqual || (isa<ConstantInt>(Val) && 1224 Val->getType()->isIntegerTy(1))) { 1225 Value *Replacement; 1226 if (IsEqual) 1227 Replacement = Val; 1228 else 1229 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 1230 !cast<ConstantInt>(Val)->getZExtValue()); 1231 1232 for (User *U : LIC->users()) { 1233 Instruction *UI = dyn_cast<Instruction>(U); 1234 if (!UI || !L->contains(UI)) 1235 continue; 1236 Worklist.push_back(UI); 1237 } 1238 1239 for (Instruction *UI : Worklist) 1240 UI->replaceUsesOfWith(LIC, Replacement); 1241 1242 SimplifyCode(Worklist, L); 1243 return; 1244 } 1245 1246 // Otherwise, we don't know the precise value of LIC, but we do know that it 1247 // is certainly NOT "Val". As such, simplify any uses in the loop that we 1248 // can. This case occurs when we unswitch switch statements. 1249 for (User *U : LIC->users()) { 1250 Instruction *UI = dyn_cast<Instruction>(U); 1251 if (!UI || !L->contains(UI)) 1252 continue; 1253 1254 Worklist.push_back(UI); 1255 1256 // TODO: We could do other simplifications, for example, turning 1257 // 'icmp eq LIC, Val' -> false. 1258 1259 // If we know that LIC is not Val, use this info to simplify code. 1260 SwitchInst *SI = dyn_cast<SwitchInst>(UI); 1261 if (!SI || !isa<ConstantInt>(Val)) continue; 1262 1263 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val)); 1264 // Default case is live for multiple values. 1265 if (DeadCase == SI->case_default()) continue; 1266 1267 // Found a dead case value. Don't remove PHI nodes in the 1268 // successor if they become single-entry, those PHI nodes may 1269 // be in the Users list. 1270 1271 BasicBlock *Switch = SI->getParent(); 1272 BasicBlock *SISucc = DeadCase.getCaseSuccessor(); 1273 BasicBlock *Latch = L->getLoopLatch(); 1274 1275 BranchesInfo.setUnswitched(SI, Val); 1276 1277 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical. 1278 // If the DeadCase successor dominates the loop latch, then the 1279 // transformation isn't safe since it will delete the sole predecessor edge 1280 // to the latch. 1281 if (Latch && DT->dominates(SISucc, Latch)) 1282 continue; 1283 1284 // FIXME: This is a hack. We need to keep the successor around 1285 // and hooked up so as to preserve the loop structure, because 1286 // trying to update it is complicated. So instead we preserve the 1287 // loop structure and put the block on a dead code path. 1288 SplitEdge(Switch, SISucc, DT, LI); 1289 // Compute the successors instead of relying on the return value 1290 // of SplitEdge, since it may have split the switch successor 1291 // after PHI nodes. 1292 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor(); 1293 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 1294 // Create an "unreachable" destination. 1295 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 1296 Switch->getParent(), 1297 OldSISucc); 1298 new UnreachableInst(Context, Abort); 1299 // Force the new case destination to branch to the "unreachable" 1300 // block while maintaining a (dead) CFG edge to the old block. 1301 NewSISucc->getTerminator()->eraseFromParent(); 1302 BranchInst::Create(Abort, OldSISucc, 1303 ConstantInt::getTrue(Context), NewSISucc); 1304 // Release the PHI operands for this edge. 1305 for (BasicBlock::iterator II = NewSISucc->begin(); 1306 PHINode *PN = dyn_cast<PHINode>(II); ++II) 1307 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 1308 UndefValue::get(PN->getType())); 1309 // Tell the domtree about the new block. We don't fully update the 1310 // domtree here -- instead we force it to do a full recomputation 1311 // after the pass is complete -- but we do need to inform it of 1312 // new blocks. 1313 DT->addNewBlock(Abort, NewSISucc); 1314 } 1315 1316 SimplifyCode(Worklist, L); 1317 } 1318 1319 /// Now that we have simplified some instructions in the loop, walk over it and 1320 /// constant prop, dce, and fold control flow where possible. Note that this is 1321 /// effectively a very simple loop-structure-aware optimizer. During processing 1322 /// of this loop, L could very well be deleted, so it must not be used. 1323 /// 1324 /// FIXME: When the loop optimizer is more mature, separate this out to a new 1325 /// pass. 1326 /// 1327 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 1328 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 1329 while (!Worklist.empty()) { 1330 Instruction *I = Worklist.back(); 1331 Worklist.pop_back(); 1332 1333 // Simple DCE. 1334 if (isInstructionTriviallyDead(I)) { 1335 DEBUG(dbgs() << "Remove dead instruction '" << *I); 1336 1337 // Add uses to the worklist, which may be dead now. 1338 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1339 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1340 Worklist.push_back(Use); 1341 LPM->deleteSimpleAnalysisValue(I, L); 1342 RemoveFromWorklist(I, Worklist); 1343 I->eraseFromParent(); 1344 ++NumSimplify; 1345 continue; 1346 } 1347 1348 // See if instruction simplification can hack this up. This is common for 1349 // things like "select false, X, Y" after unswitching made the condition be 1350 // 'false'. TODO: update the domtree properly so we can pass it here. 1351 if (Value *V = SimplifyInstruction(I, DL)) 1352 if (LI->replacementPreservesLCSSAForm(I, V)) { 1353 ReplaceUsesOfWith(I, V, Worklist, L, LPM); 1354 continue; 1355 } 1356 1357 // Special case hacks that appear commonly in unswitched code. 1358 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1359 if (BI->isUnconditional()) { 1360 // If BI's parent is the only pred of the successor, fold the two blocks 1361 // together. 1362 BasicBlock *Pred = BI->getParent(); 1363 BasicBlock *Succ = BI->getSuccessor(0); 1364 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1365 if (!SinglePred) continue; // Nothing to do. 1366 assert(SinglePred == Pred && "CFG broken"); 1367 1368 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1369 << Succ->getName() << "\n"); 1370 1371 // Resolve any single entry PHI nodes in Succ. 1372 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1373 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1374 1375 // If Succ has any successors with PHI nodes, update them to have 1376 // entries coming from Pred instead of Succ. 1377 Succ->replaceAllUsesWith(Pred); 1378 1379 // Move all of the successor contents from Succ to Pred. 1380 Pred->getInstList().splice(BI->getIterator(), Succ->getInstList(), 1381 Succ->begin(), Succ->end()); 1382 LPM->deleteSimpleAnalysisValue(BI, L); 1383 BI->eraseFromParent(); 1384 RemoveFromWorklist(BI, Worklist); 1385 1386 // Remove Succ from the loop tree. 1387 LI->removeBlock(Succ); 1388 LPM->deleteSimpleAnalysisValue(Succ, L); 1389 Succ->eraseFromParent(); 1390 ++NumSimplify; 1391 continue; 1392 } 1393 1394 continue; 1395 } 1396 } 1397 } 1398