1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===// 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 implements a simple loop unroller. It works best when loops have 11 // been canonicalized by the -indvars pass, allowing it to determine the trip 12 // counts of loops easily. 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Scalar.h" 16 #include "llvm/ADT/SetVector.h" 17 #include "llvm/Analysis/GlobalsModRef.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/CodeMetrics.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/LoopPass.h" 22 #include "llvm/Analysis/ScalarEvolution.h" 23 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 24 #include "llvm/Analysis/TargetTransformInfo.h" 25 #include "llvm/IR/DataLayout.h" 26 #include "llvm/IR/DiagnosticInfo.h" 27 #include "llvm/IR/Dominators.h" 28 #include "llvm/IR/InstVisitor.h" 29 #include "llvm/IR/IntrinsicInst.h" 30 #include "llvm/IR/Metadata.h" 31 #include "llvm/Support/CommandLine.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/raw_ostream.h" 34 #include "llvm/Transforms/Utils/UnrollLoop.h" 35 #include <climits> 36 37 using namespace llvm; 38 39 #define DEBUG_TYPE "loop-unroll" 40 41 static cl::opt<unsigned> 42 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden, 43 cl::desc("The baseline cost threshold for loop unrolling")); 44 45 static cl::opt<unsigned> UnrollPercentDynamicCostSavedThreshold( 46 "unroll-percent-dynamic-cost-saved-threshold", cl::init(20), cl::Hidden, 47 cl::desc("The percentage of estimated dynamic cost which must be saved by " 48 "unrolling to allow unrolling up to the max threshold.")); 49 50 static cl::opt<unsigned> UnrollDynamicCostSavingsDiscount( 51 "unroll-dynamic-cost-savings-discount", cl::init(2000), cl::Hidden, 52 cl::desc("This is the amount discounted from the total unroll cost when " 53 "the unrolled form has a high dynamic cost savings (triggered by " 54 "the '-unroll-perecent-dynamic-cost-saved-threshold' flag).")); 55 56 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze( 57 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden, 58 cl::desc("Don't allow loop unrolling to simulate more than this number of" 59 "iterations when checking full unroll profitability")); 60 61 static cl::opt<unsigned> 62 UnrollCount("unroll-count", cl::init(0), cl::Hidden, 63 cl::desc("Use this unroll count for all loops including those with " 64 "unroll_count pragma values, for testing purposes")); 65 66 static cl::opt<bool> 67 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden, 68 cl::desc("Allows loops to be partially unrolled until " 69 "-unroll-threshold loop size is reached.")); 70 71 static cl::opt<bool> 72 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden, 73 cl::desc("Unroll loops with run-time trip counts")); 74 75 static cl::opt<unsigned> 76 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden, 77 cl::desc("Unrolled size limit for loops with an unroll(full) or " 78 "unroll_count pragma.")); 79 80 namespace { 81 class LoopUnroll : public LoopPass { 82 public: 83 static char ID; // Pass ID, replacement for typeid 84 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) { 85 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T); 86 CurrentPercentDynamicCostSavedThreshold = 87 UnrollPercentDynamicCostSavedThreshold; 88 CurrentDynamicCostSavingsDiscount = UnrollDynamicCostSavingsDiscount; 89 CurrentCount = (C == -1) ? UnrollCount : unsigned(C); 90 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P; 91 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R; 92 93 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0); 94 UserPercentDynamicCostSavedThreshold = 95 (UnrollPercentDynamicCostSavedThreshold.getNumOccurrences() > 0); 96 UserDynamicCostSavingsDiscount = 97 (UnrollDynamicCostSavingsDiscount.getNumOccurrences() > 0); 98 UserAllowPartial = (P != -1) || 99 (UnrollAllowPartial.getNumOccurrences() > 0); 100 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0); 101 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0); 102 103 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 104 } 105 106 /// A magic value for use with the Threshold parameter to indicate 107 /// that the loop unroll should be performed regardless of how much 108 /// code expansion would result. 109 static const unsigned NoThreshold = UINT_MAX; 110 111 // Threshold to use when optsize is specified (and there is no 112 // explicit -unroll-threshold). 113 static const unsigned OptSizeUnrollThreshold = 50; 114 115 // Default unroll count for loops with run-time trip count if 116 // -unroll-count is not set 117 static const unsigned UnrollRuntimeCount = 8; 118 119 unsigned CurrentCount; 120 unsigned CurrentThreshold; 121 unsigned CurrentPercentDynamicCostSavedThreshold; 122 unsigned CurrentDynamicCostSavingsDiscount; 123 bool CurrentAllowPartial; 124 bool CurrentRuntime; 125 126 // Flags for whether the 'current' settings are user-specified. 127 bool UserCount; 128 bool UserThreshold; 129 bool UserPercentDynamicCostSavedThreshold; 130 bool UserDynamicCostSavingsDiscount; 131 bool UserAllowPartial; 132 bool UserRuntime; 133 134 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 135 136 /// This transformation requires natural loop information & requires that 137 /// loop preheaders be inserted into the CFG... 138 /// 139 void getAnalysisUsage(AnalysisUsage &AU) const override { 140 AU.addRequired<AssumptionCacheTracker>(); 141 AU.addRequired<DominatorTreeWrapperPass>(); 142 AU.addRequired<LoopInfoWrapperPass>(); 143 AU.addPreserved<LoopInfoWrapperPass>(); 144 AU.addRequiredID(LoopSimplifyID); 145 AU.addPreservedID(LoopSimplifyID); 146 AU.addRequiredID(LCSSAID); 147 AU.addPreservedID(LCSSAID); 148 AU.addRequired<ScalarEvolutionWrapperPass>(); 149 AU.addPreserved<ScalarEvolutionWrapperPass>(); 150 AU.addRequired<TargetTransformInfoWrapperPass>(); 151 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 152 // If loop unroll does not preserve dom info then LCSSA pass on next 153 // loop will receive invalid dom info. 154 // For now, recreate dom info, if loop is unrolled. 155 AU.addPreserved<DominatorTreeWrapperPass>(); 156 AU.addPreserved<GlobalsAAWrapperPass>(); 157 } 158 159 // Fill in the UnrollingPreferences parameter with values from the 160 // TargetTransformationInfo. 161 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI, 162 TargetTransformInfo::UnrollingPreferences &UP) { 163 UP.Threshold = CurrentThreshold; 164 UP.PercentDynamicCostSavedThreshold = 165 CurrentPercentDynamicCostSavedThreshold; 166 UP.DynamicCostSavingsDiscount = CurrentDynamicCostSavingsDiscount; 167 UP.OptSizeThreshold = OptSizeUnrollThreshold; 168 UP.PartialThreshold = CurrentThreshold; 169 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold; 170 UP.Count = CurrentCount; 171 UP.MaxCount = UINT_MAX; 172 UP.Partial = CurrentAllowPartial; 173 UP.Runtime = CurrentRuntime; 174 UP.AllowExpensiveTripCount = false; 175 TTI.getUnrollingPreferences(L, UP); 176 } 177 178 // Select and return an unroll count based on parameters from 179 // user, unroll preferences, unroll pragmas, or a heuristic. 180 // SetExplicitly is set to true if the unroll count is is set by 181 // the user or a pragma rather than selected heuristically. 182 unsigned 183 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll, 184 unsigned PragmaCount, 185 const TargetTransformInfo::UnrollingPreferences &UP, 186 bool &SetExplicitly); 187 188 // Select threshold values used to limit unrolling based on a 189 // total unrolled size. Parameters Threshold and PartialThreshold 190 // are set to the maximum unrolled size for fully and partially 191 // unrolled loops respectively. 192 void selectThresholds(const Loop *L, bool UsePragmaThreshold, 193 const TargetTransformInfo::UnrollingPreferences &UP, 194 unsigned &Threshold, unsigned &PartialThreshold, 195 unsigned &PercentDynamicCostSavedThreshold, 196 unsigned &DynamicCostSavingsDiscount) { 197 // Determine the current unrolling threshold. While this is 198 // normally set from UnrollThreshold, it is overridden to a 199 // smaller value if the current function is marked as 200 // optimize-for-size, and the unroll threshold was not user 201 // specified. 202 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold; 203 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold; 204 PercentDynamicCostSavedThreshold = 205 UserPercentDynamicCostSavedThreshold 206 ? CurrentPercentDynamicCostSavedThreshold 207 : UP.PercentDynamicCostSavedThreshold; 208 DynamicCostSavingsDiscount = UserDynamicCostSavingsDiscount 209 ? CurrentDynamicCostSavingsDiscount 210 : UP.DynamicCostSavingsDiscount; 211 212 if (!UserThreshold && 213 // FIXME: Use Function::optForSize(). 214 L->getHeader()->getParent()->hasFnAttribute( 215 Attribute::OptimizeForSize)) { 216 Threshold = UP.OptSizeThreshold; 217 PartialThreshold = UP.PartialOptSizeThreshold; 218 } 219 if (UsePragmaThreshold) { 220 // If the loop has an unrolling pragma, we want to be more 221 // aggressive with unrolling limits. Set thresholds to at 222 // least the PragmaTheshold value which is larger than the 223 // default limits. 224 if (Threshold != NoThreshold) 225 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold); 226 if (PartialThreshold != NoThreshold) 227 PartialThreshold = 228 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold); 229 } 230 } 231 bool canUnrollCompletely(Loop *L, unsigned Threshold, 232 unsigned PercentDynamicCostSavedThreshold, 233 unsigned DynamicCostSavingsDiscount, 234 uint64_t UnrolledCost, uint64_t RolledDynamicCost); 235 }; 236 } 237 238 char LoopUnroll::ID = 0; 239 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 240 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 241 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 242 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 243 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 244 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 245 INITIALIZE_PASS_DEPENDENCY(LCSSA) 246 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 247 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 248 249 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial, 250 int Runtime) { 251 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime); 252 } 253 254 Pass *llvm::createSimpleLoopUnrollPass() { 255 return llvm::createLoopUnrollPass(-1, -1, 0, 0); 256 } 257 258 namespace { 259 // This class is used to get an estimate of the optimization effects that we 260 // could get from complete loop unrolling. It comes from the fact that some 261 // loads might be replaced with concrete constant values and that could trigger 262 // a chain of instruction simplifications. 263 // 264 // E.g. we might have: 265 // int a[] = {0, 1, 0}; 266 // v = 0; 267 // for (i = 0; i < 3; i ++) 268 // v += b[i]*a[i]; 269 // If we completely unroll the loop, we would get: 270 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2] 271 // Which then will be simplified to: 272 // v = b[0]* 0 + b[1]* 1 + b[2]* 0 273 // And finally: 274 // v = b[1] 275 class UnrolledInstAnalyzer : private InstVisitor<UnrolledInstAnalyzer, bool> { 276 typedef InstVisitor<UnrolledInstAnalyzer, bool> Base; 277 friend class InstVisitor<UnrolledInstAnalyzer, bool>; 278 struct SimplifiedAddress { 279 Value *Base = nullptr; 280 ConstantInt *Offset = nullptr; 281 }; 282 283 public: 284 UnrolledInstAnalyzer(unsigned Iteration, 285 DenseMap<Value *, Constant *> &SimplifiedValues, 286 const Loop *L, ScalarEvolution &SE) 287 : Iteration(Iteration), SimplifiedValues(SimplifiedValues), L(L), SE(SE) { 288 IterationNumber = SE.getConstant(APInt(64, Iteration)); 289 } 290 291 // Allow access to the initial visit method. 292 using Base::visit; 293 294 private: 295 /// \brief A cache of pointer bases and constant-folded offsets corresponding 296 /// to GEP (or derived from GEP) instructions. 297 /// 298 /// In order to find the base pointer one needs to perform non-trivial 299 /// traversal of the corresponding SCEV expression, so it's good to have the 300 /// results saved. 301 DenseMap<Value *, SimplifiedAddress> SimplifiedAddresses; 302 303 /// \brief Number of currently simulated iteration. 304 /// 305 /// If an expression is ConstAddress+Constant, then the Constant is 306 /// Start + Iteration*Step, where Start and Step could be obtained from 307 /// SCEVGEPCache. 308 unsigned Iteration; 309 310 /// \brief SCEV expression corresponding to number of currently simulated 311 /// iteration. 312 const SCEV *IterationNumber; 313 314 /// \brief A Value->Constant map for keeping values that we managed to 315 /// constant-fold on the given iteration. 316 /// 317 /// While we walk the loop instructions, we build up and maintain a mapping 318 /// of simplified values specific to this iteration. The idea is to propagate 319 /// any special information we have about loads that can be replaced with 320 /// constants after complete unrolling, and account for likely simplifications 321 /// post-unrolling. 322 DenseMap<Value *, Constant *> &SimplifiedValues; 323 324 const Loop *L; 325 ScalarEvolution &SE; 326 327 /// \brief Try to simplify instruction \param I using its SCEV expression. 328 /// 329 /// The idea is that some AddRec expressions become constants, which then 330 /// could trigger folding of other instructions. However, that only happens 331 /// for expressions whose start value is also constant, which isn't always the 332 /// case. In another common and important case the start value is just some 333 /// address (i.e. SCEVUnknown) - in this case we compute the offset and save 334 /// it along with the base address instead. 335 bool simplifyInstWithSCEV(Instruction *I) { 336 if (!SE.isSCEVable(I->getType())) 337 return false; 338 339 const SCEV *S = SE.getSCEV(I); 340 if (auto *SC = dyn_cast<SCEVConstant>(S)) { 341 SimplifiedValues[I] = SC->getValue(); 342 return true; 343 } 344 345 auto *AR = dyn_cast<SCEVAddRecExpr>(S); 346 if (!AR) 347 return false; 348 349 const SCEV *ValueAtIteration = AR->evaluateAtIteration(IterationNumber, SE); 350 // Check if the AddRec expression becomes a constant. 351 if (auto *SC = dyn_cast<SCEVConstant>(ValueAtIteration)) { 352 SimplifiedValues[I] = SC->getValue(); 353 return true; 354 } 355 356 // Check if the offset from the base address becomes a constant. 357 auto *Base = dyn_cast<SCEVUnknown>(SE.getPointerBase(S)); 358 if (!Base) 359 return false; 360 auto *Offset = 361 dyn_cast<SCEVConstant>(SE.getMinusSCEV(ValueAtIteration, Base)); 362 if (!Offset) 363 return false; 364 SimplifiedAddress Address; 365 Address.Base = Base->getValue(); 366 Address.Offset = Offset->getValue(); 367 SimplifiedAddresses[I] = Address; 368 return true; 369 } 370 371 /// Base case for the instruction visitor. 372 bool visitInstruction(Instruction &I) { 373 return simplifyInstWithSCEV(&I); 374 } 375 376 /// Try to simplify binary operator I. 377 /// 378 /// TODO: Probably it's worth to hoist the code for estimating the 379 /// simplifications effects to a separate class, since we have a very similar 380 /// code in InlineCost already. 381 bool visitBinaryOperator(BinaryOperator &I) { 382 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 383 if (!isa<Constant>(LHS)) 384 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 385 LHS = SimpleLHS; 386 if (!isa<Constant>(RHS)) 387 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 388 RHS = SimpleRHS; 389 390 Value *SimpleV = nullptr; 391 const DataLayout &DL = I.getModule()->getDataLayout(); 392 if (auto FI = dyn_cast<FPMathOperator>(&I)) 393 SimpleV = 394 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags(), DL); 395 else 396 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL); 397 398 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) 399 SimplifiedValues[&I] = C; 400 401 if (SimpleV) 402 return true; 403 return Base::visitBinaryOperator(I); 404 } 405 406 /// Try to fold load I. 407 bool visitLoad(LoadInst &I) { 408 Value *AddrOp = I.getPointerOperand(); 409 410 auto AddressIt = SimplifiedAddresses.find(AddrOp); 411 if (AddressIt == SimplifiedAddresses.end()) 412 return false; 413 ConstantInt *SimplifiedAddrOp = AddressIt->second.Offset; 414 415 auto *GV = dyn_cast<GlobalVariable>(AddressIt->second.Base); 416 // We're only interested in loads that can be completely folded to a 417 // constant. 418 if (!GV || !GV->hasInitializer() || !GV->isConstant()) 419 return false; 420 421 ConstantDataSequential *CDS = 422 dyn_cast<ConstantDataSequential>(GV->getInitializer()); 423 if (!CDS) 424 return false; 425 426 int ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U; 427 assert(SimplifiedAddrOp->getValue().getActiveBits() < 64 && 428 "Unexpectedly large index value."); 429 int64_t Index = SimplifiedAddrOp->getSExtValue() / ElemSize; 430 if (Index >= CDS->getNumElements()) { 431 // FIXME: For now we conservatively ignore out of bound accesses, but 432 // we're allowed to perform the optimization in this case. 433 return false; 434 } 435 436 Constant *CV = CDS->getElementAsConstant(Index); 437 assert(CV && "Constant expected."); 438 SimplifiedValues[&I] = CV; 439 440 return true; 441 } 442 443 bool visitCastInst(CastInst &I) { 444 // Propagate constants through casts. 445 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 446 if (!COp) 447 COp = SimplifiedValues.lookup(I.getOperand(0)); 448 if (COp) 449 if (Constant *C = 450 ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) { 451 SimplifiedValues[&I] = C; 452 return true; 453 } 454 455 return Base::visitCastInst(I); 456 } 457 458 bool visitCmpInst(CmpInst &I) { 459 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 460 461 // First try to handle simplified comparisons. 462 if (!isa<Constant>(LHS)) 463 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 464 LHS = SimpleLHS; 465 if (!isa<Constant>(RHS)) 466 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 467 RHS = SimpleRHS; 468 469 if (!isa<Constant>(LHS) && !isa<Constant>(RHS)) { 470 auto SimplifiedLHS = SimplifiedAddresses.find(LHS); 471 if (SimplifiedLHS != SimplifiedAddresses.end()) { 472 auto SimplifiedRHS = SimplifiedAddresses.find(RHS); 473 if (SimplifiedRHS != SimplifiedAddresses.end()) { 474 SimplifiedAddress &LHSAddr = SimplifiedLHS->second; 475 SimplifiedAddress &RHSAddr = SimplifiedRHS->second; 476 if (LHSAddr.Base == RHSAddr.Base) { 477 LHS = LHSAddr.Offset; 478 RHS = RHSAddr.Offset; 479 } 480 } 481 } 482 } 483 484 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 485 if (Constant *CRHS = dyn_cast<Constant>(RHS)) { 486 if (Constant *C = ConstantExpr::getCompare(I.getPredicate(), CLHS, CRHS)) { 487 SimplifiedValues[&I] = C; 488 return true; 489 } 490 } 491 } 492 493 return Base::visitCmpInst(I); 494 } 495 }; 496 } // namespace 497 498 499 namespace { 500 struct EstimatedUnrollCost { 501 /// \brief The estimated cost after unrolling. 502 int UnrolledCost; 503 504 /// \brief The estimated dynamic cost of executing the instructions in the 505 /// rolled form. 506 int RolledDynamicCost; 507 }; 508 } 509 510 /// \brief Figure out if the loop is worth full unrolling. 511 /// 512 /// Complete loop unrolling can make some loads constant, and we need to know 513 /// if that would expose any further optimization opportunities. This routine 514 /// estimates this optimization. It computes cost of unrolled loop 515 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By 516 /// dynamic cost we mean that we won't count costs of blocks that are known not 517 /// to be executed (i.e. if we have a branch in the loop and we know that at the 518 /// given iteration its condition would be resolved to true, we won't add up the 519 /// cost of the 'false'-block). 520 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If 521 /// the analysis failed (no benefits expected from the unrolling, or the loop is 522 /// too big to analyze), the returned value is None. 523 static Optional<EstimatedUnrollCost> 524 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, DominatorTree &DT, 525 ScalarEvolution &SE, const TargetTransformInfo &TTI, 526 int MaxUnrolledLoopSize) { 527 // We want to be able to scale offsets by the trip count and add more offsets 528 // to them without checking for overflows, and we already don't want to 529 // analyze *massive* trip counts, so we force the max to be reasonably small. 530 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) && 531 "The unroll iterations max is too large!"); 532 533 // Don't simulate loops with a big or unknown tripcount 534 if (!UnrollMaxIterationsCountToAnalyze || !TripCount || 535 TripCount > UnrollMaxIterationsCountToAnalyze) 536 return None; 537 538 SmallSetVector<BasicBlock *, 16> BBWorklist; 539 DenseMap<Value *, Constant *> SimplifiedValues; 540 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues; 541 542 // The estimated cost of the unrolled form of the loop. We try to estimate 543 // this by simplifying as much as we can while computing the estimate. 544 int UnrolledCost = 0; 545 // We also track the estimated dynamic (that is, actually executed) cost in 546 // the rolled form. This helps identify cases when the savings from unrolling 547 // aren't just exposing dead control flows, but actual reduced dynamic 548 // instructions due to the simplifications which we expect to occur after 549 // unrolling. 550 int RolledDynamicCost = 0; 551 552 // Ensure that we don't violate the loop structure invariants relied on by 553 // this analysis. 554 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first."); 555 assert(L->isLCSSAForm(DT) && 556 "Must have loops in LCSSA form to track live-out values."); 557 558 DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n"); 559 560 // Simulate execution of each iteration of the loop counting instructions, 561 // which would be simplified. 562 // Since the same load will take different values on different iterations, 563 // we literally have to go through all loop's iterations. 564 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) { 565 DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n"); 566 567 // Prepare for the iteration by collecting any simplified entry or backedge 568 // inputs. 569 for (Instruction &I : *L->getHeader()) { 570 auto *PHI = dyn_cast<PHINode>(&I); 571 if (!PHI) 572 break; 573 574 // The loop header PHI nodes must have exactly two input: one from the 575 // loop preheader and one from the loop latch. 576 assert( 577 PHI->getNumIncomingValues() == 2 && 578 "Must have an incoming value only for the preheader and the latch."); 579 580 Value *V = PHI->getIncomingValueForBlock( 581 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch()); 582 Constant *C = dyn_cast<Constant>(V); 583 if (Iteration != 0 && !C) 584 C = SimplifiedValues.lookup(V); 585 if (C) 586 SimplifiedInputValues.push_back({PHI, C}); 587 } 588 589 // Now clear and re-populate the map for the next iteration. 590 SimplifiedValues.clear(); 591 while (!SimplifiedInputValues.empty()) 592 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val()); 593 594 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, L, SE); 595 596 BBWorklist.clear(); 597 BBWorklist.insert(L->getHeader()); 598 // Note that we *must not* cache the size, this loop grows the worklist. 599 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) { 600 BasicBlock *BB = BBWorklist[Idx]; 601 602 // Visit all instructions in the given basic block and try to simplify 603 // it. We don't change the actual IR, just count optimization 604 // opportunities. 605 for (Instruction &I : *BB) { 606 int InstCost = TTI.getUserCost(&I); 607 608 // Visit the instruction to analyze its loop cost after unrolling, 609 // and if the visitor returns false, include this instruction in the 610 // unrolled cost. 611 if (!Analyzer.visit(I)) 612 UnrolledCost += InstCost; 613 else { 614 DEBUG(dbgs() << " " << I 615 << " would be simplified if loop is unrolled.\n"); 616 (void)0; 617 } 618 619 // Also track this instructions expected cost when executing the rolled 620 // loop form. 621 RolledDynamicCost += InstCost; 622 623 // If unrolled body turns out to be too big, bail out. 624 if (UnrolledCost > MaxUnrolledLoopSize) { 625 DEBUG(dbgs() << " Exceeded threshold.. exiting.\n" 626 << " UnrolledCost: " << UnrolledCost 627 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize 628 << "\n"); 629 return None; 630 } 631 } 632 633 TerminatorInst *TI = BB->getTerminator(); 634 635 // Add in the live successors by first checking whether we have terminator 636 // that may be simplified based on the values simplified by this call. 637 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 638 if (BI->isConditional()) { 639 if (Constant *SimpleCond = 640 SimplifiedValues.lookup(BI->getCondition())) { 641 BasicBlock *Succ = nullptr; 642 // Just take the first successor if condition is undef 643 if (isa<UndefValue>(SimpleCond)) 644 Succ = BI->getSuccessor(0); 645 else 646 Succ = BI->getSuccessor( 647 cast<ConstantInt>(SimpleCond)->isZero() ? 1 : 0); 648 if (L->contains(Succ)) 649 BBWorklist.insert(Succ); 650 continue; 651 } 652 } 653 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 654 if (Constant *SimpleCond = 655 SimplifiedValues.lookup(SI->getCondition())) { 656 BasicBlock *Succ = nullptr; 657 // Just take the first successor if condition is undef 658 if (isa<UndefValue>(SimpleCond)) 659 Succ = SI->getSuccessor(0); 660 else 661 Succ = SI->findCaseValue(cast<ConstantInt>(SimpleCond)) 662 .getCaseSuccessor(); 663 if (L->contains(Succ)) 664 BBWorklist.insert(Succ); 665 continue; 666 } 667 } 668 669 // Add BB's successors to the worklist. 670 for (BasicBlock *Succ : successors(BB)) 671 if (L->contains(Succ)) 672 BBWorklist.insert(Succ); 673 } 674 675 // If we found no optimization opportunities on the first iteration, we 676 // won't find them on later ones too. 677 if (UnrolledCost == RolledDynamicCost) { 678 DEBUG(dbgs() << " No opportunities found.. exiting.\n" 679 << " UnrolledCost: " << UnrolledCost << "\n"); 680 return None; 681 } 682 } 683 DEBUG(dbgs() << "Analysis finished:\n" 684 << "UnrolledCost: " << UnrolledCost << ", " 685 << "RolledDynamicCost: " << RolledDynamicCost << "\n"); 686 return {{UnrolledCost, RolledDynamicCost}}; 687 } 688 689 /// ApproximateLoopSize - Approximate the size of the loop. 690 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, 691 bool &NotDuplicatable, 692 const TargetTransformInfo &TTI, 693 AssumptionCache *AC) { 694 SmallPtrSet<const Value *, 32> EphValues; 695 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 696 697 CodeMetrics Metrics; 698 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 699 I != E; ++I) 700 Metrics.analyzeBasicBlock(*I, TTI, EphValues); 701 NumCalls = Metrics.NumInlineCandidates; 702 NotDuplicatable = Metrics.notDuplicatable; 703 704 unsigned LoopSize = Metrics.NumInsts; 705 706 // Don't allow an estimate of size zero. This would allows unrolling of loops 707 // with huge iteration counts, which is a compile time problem even if it's 708 // not a problem for code quality. Also, the code using this size may assume 709 // that each loop has at least three instructions (likely a conditional 710 // branch, a comparison feeding that branch, and some kind of loop increment 711 // feeding that comparison instruction). 712 LoopSize = std::max(LoopSize, 3u); 713 714 return LoopSize; 715 } 716 717 // Returns the loop hint metadata node with the given name (for example, 718 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is 719 // returned. 720 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) { 721 if (MDNode *LoopID = L->getLoopID()) 722 return GetUnrollMetadata(LoopID, Name); 723 return nullptr; 724 } 725 726 // Returns true if the loop has an unroll(full) pragma. 727 static bool HasUnrollFullPragma(const Loop *L) { 728 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full"); 729 } 730 731 // Returns true if the loop has an unroll(enable) pragma. This metadata is used 732 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives. 733 static bool HasUnrollEnablePragma(const Loop *L) { 734 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable"); 735 } 736 737 // Returns true if the loop has an unroll(disable) pragma. 738 static bool HasUnrollDisablePragma(const Loop *L) { 739 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable"); 740 } 741 742 // Returns true if the loop has an runtime unroll(disable) pragma. 743 static bool HasRuntimeUnrollDisablePragma(const Loop *L) { 744 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable"); 745 } 746 747 // If loop has an unroll_count pragma return the (necessarily 748 // positive) value from the pragma. Otherwise return 0. 749 static unsigned UnrollCountPragmaValue(const Loop *L) { 750 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count"); 751 if (MD) { 752 assert(MD->getNumOperands() == 2 && 753 "Unroll count hint metadata should have two operands."); 754 unsigned Count = 755 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 756 assert(Count >= 1 && "Unroll count must be positive."); 757 return Count; 758 } 759 return 0; 760 } 761 762 // Remove existing unroll metadata and add unroll disable metadata to 763 // indicate the loop has already been unrolled. This prevents a loop 764 // from being unrolled more than is directed by a pragma if the loop 765 // unrolling pass is run more than once (which it generally is). 766 static void SetLoopAlreadyUnrolled(Loop *L) { 767 MDNode *LoopID = L->getLoopID(); 768 if (!LoopID) return; 769 770 // First remove any existing loop unrolling metadata. 771 SmallVector<Metadata *, 4> MDs; 772 // Reserve first location for self reference to the LoopID metadata node. 773 MDs.push_back(nullptr); 774 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 775 bool IsUnrollMetadata = false; 776 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 777 if (MD) { 778 const MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 779 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll."); 780 } 781 if (!IsUnrollMetadata) 782 MDs.push_back(LoopID->getOperand(i)); 783 } 784 785 // Add unroll(disable) metadata to disable future unrolling. 786 LLVMContext &Context = L->getHeader()->getContext(); 787 SmallVector<Metadata *, 1> DisableOperands; 788 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable")); 789 MDNode *DisableNode = MDNode::get(Context, DisableOperands); 790 MDs.push_back(DisableNode); 791 792 MDNode *NewLoopID = MDNode::get(Context, MDs); 793 // Set operand 0 to refer to the loop id itself. 794 NewLoopID->replaceOperandWith(0, NewLoopID); 795 L->setLoopID(NewLoopID); 796 } 797 798 bool LoopUnroll::canUnrollCompletely(Loop *L, unsigned Threshold, 799 unsigned PercentDynamicCostSavedThreshold, 800 unsigned DynamicCostSavingsDiscount, 801 uint64_t UnrolledCost, 802 uint64_t RolledDynamicCost) { 803 804 if (Threshold == NoThreshold) { 805 DEBUG(dbgs() << " Can fully unroll, because no threshold is set.\n"); 806 return true; 807 } 808 809 if (UnrolledCost <= Threshold) { 810 DEBUG(dbgs() << " Can fully unroll, because unrolled cost: " 811 << UnrolledCost << "<" << Threshold << "\n"); 812 return true; 813 } 814 815 assert(UnrolledCost && "UnrolledCost can't be 0 at this point."); 816 assert(RolledDynamicCost >= UnrolledCost && 817 "Cannot have a higher unrolled cost than a rolled cost!"); 818 819 // Compute the percentage of the dynamic cost in the rolled form that is 820 // saved when unrolled. If unrolling dramatically reduces the estimated 821 // dynamic cost of the loop, we use a higher threshold to allow more 822 // unrolling. 823 unsigned PercentDynamicCostSaved = 824 (uint64_t)(RolledDynamicCost - UnrolledCost) * 100ull / RolledDynamicCost; 825 826 if (PercentDynamicCostSaved >= PercentDynamicCostSavedThreshold && 827 (int64_t)UnrolledCost - (int64_t)DynamicCostSavingsDiscount <= 828 (int64_t)Threshold) { 829 DEBUG(dbgs() << " Can fully unroll, because unrolling will reduce the " 830 "expected dynamic cost by " << PercentDynamicCostSaved 831 << "% (threshold: " << PercentDynamicCostSavedThreshold 832 << "%)\n" 833 << " and the unrolled cost (" << UnrolledCost 834 << ") is less than the max threshold (" 835 << DynamicCostSavingsDiscount << ").\n"); 836 return true; 837 } 838 839 DEBUG(dbgs() << " Too large to fully unroll:\n"); 840 DEBUG(dbgs() << " Threshold: " << Threshold << "\n"); 841 DEBUG(dbgs() << " Max threshold: " << DynamicCostSavingsDiscount << "\n"); 842 DEBUG(dbgs() << " Percent cost saved threshold: " 843 << PercentDynamicCostSavedThreshold << "%\n"); 844 DEBUG(dbgs() << " Unrolled cost: " << UnrolledCost << "\n"); 845 DEBUG(dbgs() << " Rolled dynamic cost: " << RolledDynamicCost << "\n"); 846 DEBUG(dbgs() << " Percent cost saved: " << PercentDynamicCostSaved 847 << "\n"); 848 return false; 849 } 850 851 unsigned LoopUnroll::selectUnrollCount( 852 const Loop *L, unsigned TripCount, bool PragmaFullUnroll, 853 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP, 854 bool &SetExplicitly) { 855 SetExplicitly = true; 856 857 // User-specified count (either as a command-line option or 858 // constructor parameter) has highest precedence. 859 unsigned Count = UserCount ? CurrentCount : 0; 860 861 // If there is no user-specified count, unroll pragmas have the next 862 // highest precedence. 863 if (Count == 0) { 864 if (PragmaCount) { 865 Count = PragmaCount; 866 } else if (PragmaFullUnroll) { 867 Count = TripCount; 868 } 869 } 870 871 if (Count == 0) 872 Count = UP.Count; 873 874 if (Count == 0) { 875 SetExplicitly = false; 876 if (TripCount == 0) 877 // Runtime trip count. 878 Count = UnrollRuntimeCount; 879 else 880 // Conservative heuristic: if we know the trip count, see if we can 881 // completely unroll (subject to the threshold, checked below); otherwise 882 // try to find greatest modulo of the trip count which is still under 883 // threshold value. 884 Count = TripCount; 885 } 886 if (TripCount && Count > TripCount) 887 return TripCount; 888 return Count; 889 } 890 891 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 892 if (skipOptnoneFunction(L)) 893 return false; 894 895 Function &F = *L->getHeader()->getParent(); 896 897 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 898 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 899 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 900 const TargetTransformInfo &TTI = 901 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 902 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 903 904 BasicBlock *Header = L->getHeader(); 905 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 906 << "] Loop %" << Header->getName() << "\n"); 907 908 if (HasUnrollDisablePragma(L)) { 909 return false; 910 } 911 bool PragmaFullUnroll = HasUnrollFullPragma(L); 912 bool PragmaEnableUnroll = HasUnrollEnablePragma(L); 913 unsigned PragmaCount = UnrollCountPragmaValue(L); 914 bool HasPragma = PragmaFullUnroll || PragmaEnableUnroll || PragmaCount > 0; 915 916 TargetTransformInfo::UnrollingPreferences UP; 917 getUnrollingPreferences(L, TTI, UP); 918 919 // Find trip count and trip multiple if count is not available 920 unsigned TripCount = 0; 921 unsigned TripMultiple = 1; 922 // If there are multiple exiting blocks but one of them is the latch, use the 923 // latch for the trip count estimation. Otherwise insist on a single exiting 924 // block for the trip count estimation. 925 BasicBlock *ExitingBlock = L->getLoopLatch(); 926 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock)) 927 ExitingBlock = L->getExitingBlock(); 928 if (ExitingBlock) { 929 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock); 930 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock); 931 } 932 933 // Select an initial unroll count. This may be reduced later based 934 // on size thresholds. 935 bool CountSetExplicitly; 936 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll, 937 PragmaCount, UP, CountSetExplicitly); 938 939 unsigned NumInlineCandidates; 940 bool notDuplicatable; 941 unsigned LoopSize = 942 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC); 943 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 944 945 // When computing the unrolled size, note that the conditional branch on the 946 // backedge and the comparison feeding it are not replicated like the rest of 947 // the loop body (which is why 2 is subtracted). 948 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2; 949 if (notDuplicatable) { 950 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable" 951 << " instructions.\n"); 952 return false; 953 } 954 if (NumInlineCandidates != 0) { 955 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 956 return false; 957 } 958 959 unsigned Threshold, PartialThreshold; 960 unsigned PercentDynamicCostSavedThreshold; 961 unsigned DynamicCostSavingsDiscount; 962 // Only use the high pragma threshold when we have a target unroll factor such 963 // as with "#pragma unroll N" or a pragma indicating full unrolling and the 964 // trip count is known. Otherwise we rely on the standard threshold to 965 // heuristically select a reasonable unroll count. 966 bool UsePragmaThreshold = 967 PragmaCount > 0 || 968 ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount != 0); 969 970 selectThresholds(L, UsePragmaThreshold, UP, Threshold, PartialThreshold, 971 PercentDynamicCostSavedThreshold, 972 DynamicCostSavingsDiscount); 973 974 // Given Count, TripCount and thresholds determine the type of 975 // unrolling which is to be performed. 976 enum { Full = 0, Partial = 1, Runtime = 2 }; 977 int Unrolling; 978 if (TripCount && Count == TripCount) { 979 Unrolling = Partial; 980 // If the loop is really small, we don't need to run an expensive analysis. 981 if (canUnrollCompletely(L, Threshold, 100, DynamicCostSavingsDiscount, 982 UnrolledSize, UnrolledSize)) { 983 Unrolling = Full; 984 } else { 985 // The loop isn't that small, but we still can fully unroll it if that 986 // helps to remove a significant number of instructions. 987 // To check that, run additional analysis on the loop. 988 if (Optional<EstimatedUnrollCost> Cost = 989 analyzeLoopUnrollCost(L, TripCount, DT, *SE, TTI, 990 Threshold + DynamicCostSavingsDiscount)) 991 if (canUnrollCompletely(L, Threshold, PercentDynamicCostSavedThreshold, 992 DynamicCostSavingsDiscount, Cost->UnrolledCost, 993 Cost->RolledDynamicCost)) { 994 Unrolling = Full; 995 } 996 } 997 } else if (TripCount && Count < TripCount) { 998 Unrolling = Partial; 999 } else { 1000 Unrolling = Runtime; 1001 } 1002 1003 // Reduce count based on the type of unrolling and the threshold values. 1004 unsigned OriginalCount = Count; 1005 bool AllowRuntime = PragmaEnableUnroll || (PragmaCount > 0) || 1006 (UserRuntime ? CurrentRuntime : UP.Runtime); 1007 // Don't unroll a runtime trip count loop with unroll full pragma. 1008 if (HasRuntimeUnrollDisablePragma(L) || PragmaFullUnroll) { 1009 AllowRuntime = false; 1010 } 1011 if (Unrolling == Partial) { 1012 bool AllowPartial = PragmaEnableUnroll || 1013 (UserAllowPartial ? CurrentAllowPartial : UP.Partial); 1014 if (!AllowPartial && !CountSetExplicitly) { 1015 DEBUG(dbgs() << " will not try to unroll partially because " 1016 << "-unroll-allow-partial not given\n"); 1017 return false; 1018 } 1019 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) { 1020 // Reduce unroll count to be modulo of TripCount for partial unrolling. 1021 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2); 1022 while (Count != 0 && TripCount % Count != 0) 1023 Count--; 1024 } 1025 } else if (Unrolling == Runtime) { 1026 if (!AllowRuntime && !CountSetExplicitly) { 1027 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count " 1028 << "-unroll-runtime not given\n"); 1029 return false; 1030 } 1031 // Reduce unroll count to be the largest power-of-two factor of 1032 // the original count which satisfies the threshold limit. 1033 while (Count != 0 && UnrolledSize > PartialThreshold) { 1034 Count >>= 1; 1035 UnrolledSize = (LoopSize-2) * Count + 2; 1036 } 1037 if (Count > UP.MaxCount) 1038 Count = UP.MaxCount; 1039 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 1040 } 1041 1042 if (HasPragma) { 1043 if (PragmaCount != 0) 1044 // If loop has an unroll count pragma mark loop as unrolled to prevent 1045 // unrolling beyond that requested by the pragma. 1046 SetLoopAlreadyUnrolled(L); 1047 1048 // Emit optimization remarks if we are unable to unroll the loop 1049 // as directed by a pragma. 1050 DebugLoc LoopLoc = L->getStartLoc(); 1051 Function *F = Header->getParent(); 1052 LLVMContext &Ctx = F->getContext(); 1053 if ((PragmaCount > 0) && Count != OriginalCount) { 1054 emitOptimizationRemarkMissed( 1055 Ctx, DEBUG_TYPE, *F, LoopLoc, 1056 "Unable to unroll loop the number of times directed by " 1057 "unroll_count pragma because unrolled size is too large."); 1058 } else if (PragmaFullUnroll && !TripCount) { 1059 emitOptimizationRemarkMissed( 1060 Ctx, DEBUG_TYPE, *F, LoopLoc, 1061 "Unable to fully unroll loop as directed by unroll(full) pragma " 1062 "because loop has a runtime trip count."); 1063 } else if (PragmaEnableUnroll && Count != TripCount && Count < 2) { 1064 emitOptimizationRemarkMissed( 1065 Ctx, DEBUG_TYPE, *F, LoopLoc, 1066 "Unable to unroll loop as directed by unroll(enable) pragma because " 1067 "unrolled size is too large."); 1068 } else if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount && 1069 Count != TripCount) { 1070 emitOptimizationRemarkMissed( 1071 Ctx, DEBUG_TYPE, *F, LoopLoc, 1072 "Unable to fully unroll loop as directed by unroll pragma because " 1073 "unrolled size is too large."); 1074 } 1075 } 1076 1077 if (Unrolling != Full && Count < 2) { 1078 // Partial unrolling by 1 is a nop. For full unrolling, a factor 1079 // of 1 makes sense because loop control can be eliminated. 1080 return false; 1081 } 1082 1083 // Unroll the loop. 1084 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount, 1085 TripMultiple, LI, this, &LPM, &AC)) 1086 return false; 1087 1088 return true; 1089 } 1090