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/AssumptionCache.h" 18 #include "llvm/Analysis/CodeMetrics.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/LoopPass.h" 21 #include "llvm/Analysis/ScalarEvolution.h" 22 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 23 #include "llvm/Analysis/TargetTransformInfo.h" 24 #include "llvm/IR/DataLayout.h" 25 #include "llvm/IR/DiagnosticInfo.h" 26 #include "llvm/IR/Dominators.h" 27 #include "llvm/IR/InstVisitor.h" 28 #include "llvm/IR/IntrinsicInst.h" 29 #include "llvm/IR/Metadata.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/Debug.h" 32 #include "llvm/Support/raw_ostream.h" 33 #include "llvm/Transforms/Utils/UnrollLoop.h" 34 #include <climits> 35 36 using namespace llvm; 37 38 #define DEBUG_TYPE "loop-unroll" 39 40 static cl::opt<unsigned> 41 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden, 42 cl::desc("The cut-off point for automatic loop unrolling")); 43 44 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze( 45 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden, 46 cl::desc("Don't allow loop unrolling to simulate more than this number of" 47 "iterations when checking full unroll profitability")); 48 49 static cl::opt<unsigned> UnrollMinPercentOfOptimized( 50 "unroll-percent-of-optimized-for-complete-unroll", cl::init(20), cl::Hidden, 51 cl::desc("If complete unrolling could trigger further optimizations, and, " 52 "by that, remove the given percent of instructions, perform the " 53 "complete unroll even if it's beyond the threshold")); 54 55 static cl::opt<unsigned> UnrollAbsoluteThreshold( 56 "unroll-absolute-threshold", cl::init(2000), cl::Hidden, 57 cl::desc("Don't unroll if the unrolled size is bigger than this threshold," 58 " even if we can remove big portion of instructions later.")); 59 60 static cl::opt<unsigned> 61 UnrollCount("unroll-count", cl::init(0), cl::Hidden, 62 cl::desc("Use this unroll count for all loops including those with " 63 "unroll_count pragma values, for testing purposes")); 64 65 static cl::opt<bool> 66 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden, 67 cl::desc("Allows loops to be partially unrolled until " 68 "-unroll-threshold loop size is reached.")); 69 70 static cl::opt<bool> 71 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden, 72 cl::desc("Unroll loops with run-time trip counts")); 73 74 static cl::opt<unsigned> 75 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden, 76 cl::desc("Unrolled size limit for loops with an unroll(full) or " 77 "unroll_count pragma.")); 78 79 namespace { 80 class LoopUnroll : public LoopPass { 81 public: 82 static char ID; // Pass ID, replacement for typeid 83 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) { 84 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T); 85 CurrentAbsoluteThreshold = UnrollAbsoluteThreshold; 86 CurrentMinPercentOfOptimized = UnrollMinPercentOfOptimized; 87 CurrentCount = (C == -1) ? UnrollCount : unsigned(C); 88 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P; 89 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R; 90 91 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0); 92 UserAbsoluteThreshold = (UnrollAbsoluteThreshold.getNumOccurrences() > 0); 93 UserPercentOfOptimized = 94 (UnrollMinPercentOfOptimized.getNumOccurrences() > 0); 95 UserAllowPartial = (P != -1) || 96 (UnrollAllowPartial.getNumOccurrences() > 0); 97 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0); 98 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0); 99 100 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 101 } 102 103 /// A magic value for use with the Threshold parameter to indicate 104 /// that the loop unroll should be performed regardless of how much 105 /// code expansion would result. 106 static const unsigned NoThreshold = UINT_MAX; 107 108 // Threshold to use when optsize is specified (and there is no 109 // explicit -unroll-threshold). 110 static const unsigned OptSizeUnrollThreshold = 50; 111 112 // Default unroll count for loops with run-time trip count if 113 // -unroll-count is not set 114 static const unsigned UnrollRuntimeCount = 8; 115 116 unsigned CurrentCount; 117 unsigned CurrentThreshold; 118 unsigned CurrentAbsoluteThreshold; 119 unsigned CurrentMinPercentOfOptimized; 120 bool CurrentAllowPartial; 121 bool CurrentRuntime; 122 bool UserCount; // CurrentCount is user-specified. 123 bool UserThreshold; // CurrentThreshold is user-specified. 124 bool UserAbsoluteThreshold; // CurrentAbsoluteThreshold is 125 // user-specified. 126 bool UserPercentOfOptimized; // CurrentMinPercentOfOptimized is 127 // user-specified. 128 bool UserAllowPartial; // CurrentAllowPartial is user-specified. 129 bool UserRuntime; // CurrentRuntime is user-specified. 130 131 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 132 133 /// This transformation requires natural loop information & requires that 134 /// loop preheaders be inserted into the CFG... 135 /// 136 void getAnalysisUsage(AnalysisUsage &AU) const override { 137 AU.addRequired<AssumptionCacheTracker>(); 138 AU.addRequired<LoopInfoWrapperPass>(); 139 AU.addPreserved<LoopInfoWrapperPass>(); 140 AU.addRequiredID(LoopSimplifyID); 141 AU.addPreservedID(LoopSimplifyID); 142 AU.addRequiredID(LCSSAID); 143 AU.addPreservedID(LCSSAID); 144 AU.addRequired<ScalarEvolution>(); 145 AU.addPreserved<ScalarEvolution>(); 146 AU.addRequired<TargetTransformInfoWrapperPass>(); 147 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 148 // If loop unroll does not preserve dom info then LCSSA pass on next 149 // loop will receive invalid dom info. 150 // For now, recreate dom info, if loop is unrolled. 151 AU.addPreserved<DominatorTreeWrapperPass>(); 152 } 153 154 // Fill in the UnrollingPreferences parameter with values from the 155 // TargetTransformationInfo. 156 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI, 157 TargetTransformInfo::UnrollingPreferences &UP) { 158 UP.Threshold = CurrentThreshold; 159 UP.AbsoluteThreshold = CurrentAbsoluteThreshold; 160 UP.MinPercentOfOptimized = CurrentMinPercentOfOptimized; 161 UP.OptSizeThreshold = OptSizeUnrollThreshold; 162 UP.PartialThreshold = CurrentThreshold; 163 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold; 164 UP.Count = CurrentCount; 165 UP.MaxCount = UINT_MAX; 166 UP.Partial = CurrentAllowPartial; 167 UP.Runtime = CurrentRuntime; 168 UP.AllowExpensiveTripCount = false; 169 TTI.getUnrollingPreferences(L, UP); 170 } 171 172 // Select and return an unroll count based on parameters from 173 // user, unroll preferences, unroll pragmas, or a heuristic. 174 // SetExplicitly is set to true if the unroll count is is set by 175 // the user or a pragma rather than selected heuristically. 176 unsigned 177 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll, 178 unsigned PragmaCount, 179 const TargetTransformInfo::UnrollingPreferences &UP, 180 bool &SetExplicitly); 181 182 // Select threshold values used to limit unrolling based on a 183 // total unrolled size. Parameters Threshold and PartialThreshold 184 // are set to the maximum unrolled size for fully and partially 185 // unrolled loops respectively. 186 void selectThresholds(const Loop *L, bool HasPragma, 187 const TargetTransformInfo::UnrollingPreferences &UP, 188 unsigned &Threshold, unsigned &PartialThreshold, 189 unsigned NumberOfOptimizedInstructions) { 190 // Determine the current unrolling threshold. While this is 191 // normally set from UnrollThreshold, it is overridden to a 192 // smaller value if the current function is marked as 193 // optimize-for-size, and the unroll threshold was not user 194 // specified. 195 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold; 196 197 // If we are allowed to completely unroll if we can remove M% of 198 // instructions, and we know that with complete unrolling we'll be able 199 // to kill N instructions, then we can afford to completely unroll loops 200 // with unrolled size up to N*100/M. 201 // Adjust the threshold according to that: 202 unsigned PercentOfOptimizedForCompleteUnroll = 203 UserPercentOfOptimized ? CurrentMinPercentOfOptimized 204 : UP.MinPercentOfOptimized; 205 unsigned AbsoluteThreshold = UserAbsoluteThreshold 206 ? CurrentAbsoluteThreshold 207 : UP.AbsoluteThreshold; 208 if (PercentOfOptimizedForCompleteUnroll) 209 Threshold = std::max<unsigned>(Threshold, 210 NumberOfOptimizedInstructions * 100 / 211 PercentOfOptimizedForCompleteUnroll); 212 // But don't allow unrolling loops bigger than absolute threshold. 213 Threshold = std::min<unsigned>(Threshold, AbsoluteThreshold); 214 215 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold; 216 if (!UserThreshold && 217 L->getHeader()->getParent()->hasFnAttribute( 218 Attribute::OptimizeForSize)) { 219 Threshold = UP.OptSizeThreshold; 220 PartialThreshold = UP.PartialOptSizeThreshold; 221 } 222 if (HasPragma) { 223 // If the loop has an unrolling pragma, we want to be more 224 // aggressive with unrolling limits. Set thresholds to at 225 // least the PragmaTheshold value which is larger than the 226 // default limits. 227 if (Threshold != NoThreshold) 228 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold); 229 if (PartialThreshold != NoThreshold) 230 PartialThreshold = 231 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold); 232 } 233 } 234 }; 235 } 236 237 char LoopUnroll::ID = 0; 238 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 239 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 240 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 241 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 242 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 243 INITIALIZE_PASS_DEPENDENCY(LCSSA) 244 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 245 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 246 247 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial, 248 int Runtime) { 249 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime); 250 } 251 252 Pass *llvm::createSimpleLoopUnrollPass() { 253 return llvm::createLoopUnrollPass(-1, -1, 0, 0); 254 } 255 256 static bool isLoadFromConstantInitializer(Value *V) { 257 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 258 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 259 return GV->getInitializer(); 260 return false; 261 } 262 263 namespace { 264 struct FindConstantPointers { 265 bool LoadCanBeConstantFolded; 266 bool IndexIsConstant; 267 APInt Step; 268 APInt StartValue; 269 Value *BaseAddress; 270 const Loop *L; 271 ScalarEvolution &SE; 272 FindConstantPointers(const Loop *loop, ScalarEvolution &SE) 273 : LoadCanBeConstantFolded(true), IndexIsConstant(true), L(loop), SE(SE) {} 274 275 bool follow(const SCEV *S) { 276 if (const SCEVUnknown *SC = dyn_cast<SCEVUnknown>(S)) { 277 // We've reached the leaf node of SCEV, it's most probably just a 278 // variable. Now it's time to see if it corresponds to a global constant 279 // global (in which case we can eliminate the load), or not. 280 BaseAddress = SC->getValue(); 281 LoadCanBeConstantFolded = 282 IndexIsConstant && isLoadFromConstantInitializer(BaseAddress); 283 return false; 284 } 285 if (isa<SCEVConstant>(S)) 286 return true; 287 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 288 // If the current SCEV expression is AddRec, and its loop isn't the loop 289 // we are about to unroll, then we won't get a constant address after 290 // unrolling, and thus, won't be able to eliminate the load. 291 if (AR->getLoop() != L) 292 return IndexIsConstant = false; 293 // If the step isn't constant, we won't get constant addresses in unrolled 294 // version. Bail out. 295 if (const SCEVConstant *StepSE = 296 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE))) 297 Step = StepSE->getValue()->getValue(); 298 else 299 return IndexIsConstant = false; 300 301 return IndexIsConstant; 302 } 303 // If Result is true, continue traversal. 304 // Otherwise, we have found something that prevents us from (possible) load 305 // elimination. 306 return IndexIsConstant; 307 } 308 bool isDone() const { return !IndexIsConstant; } 309 }; 310 311 // This class is used to get an estimate of the optimization effects that we 312 // could get from complete loop unrolling. It comes from the fact that some 313 // loads might be replaced with concrete constant values and that could trigger 314 // a chain of instruction simplifications. 315 // 316 // E.g. we might have: 317 // int a[] = {0, 1, 0}; 318 // v = 0; 319 // for (i = 0; i < 3; i ++) 320 // v += b[i]*a[i]; 321 // If we completely unroll the loop, we would get: 322 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2] 323 // Which then will be simplified to: 324 // v = b[0]* 0 + b[1]* 1 + b[2]* 0 325 // And finally: 326 // v = b[1] 327 class UnrollAnalyzer : public InstVisitor<UnrollAnalyzer, bool> { 328 typedef InstVisitor<UnrollAnalyzer, bool> Base; 329 friend class InstVisitor<UnrollAnalyzer, bool>; 330 331 const Loop *L; 332 unsigned TripCount; 333 ScalarEvolution &SE; 334 const TargetTransformInfo &TTI; 335 336 DenseMap<Value *, Constant *> SimplifiedValues; 337 DenseMap<LoadInst *, Value *> LoadBaseAddresses; 338 SmallPtrSet<Instruction *, 32> CountedInstructions; 339 340 /// \brief Count the number of optimized instructions. 341 unsigned NumberOfOptimizedInstructions; 342 343 // Provide base case for our instruction visit. 344 bool visitInstruction(Instruction &I) { return false; }; 345 // TODO: We should also visit ICmp, FCmp, GetElementPtr, Trunc, ZExt, SExt, 346 // FPTrunc, FPExt, FPToUI, FPToSI, UIToFP, SIToFP, BitCast, Select, 347 // ExtractElement, InsertElement, ShuffleVector, ExtractValue, InsertValue. 348 // 349 // Probaly it's worth to hoist the code for estimating the simplifications 350 // effects to a separate class, since we have a very similar code in 351 // InlineCost already. 352 bool visitBinaryOperator(BinaryOperator &I) { 353 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 354 if (!isa<Constant>(LHS)) 355 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 356 LHS = SimpleLHS; 357 if (!isa<Constant>(RHS)) 358 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 359 RHS = SimpleRHS; 360 Value *SimpleV = nullptr; 361 const DataLayout &DL = I.getModule()->getDataLayout(); 362 if (auto FI = dyn_cast<FPMathOperator>(&I)) 363 SimpleV = 364 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags(), DL); 365 else 366 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL); 367 368 if (SimpleV && CountedInstructions.insert(&I).second) 369 NumberOfOptimizedInstructions += TTI.getUserCost(&I); 370 371 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) { 372 SimplifiedValues[&I] = C; 373 return true; 374 } 375 return false; 376 } 377 378 Constant *computeLoadValue(LoadInst *LI, unsigned Iteration) { 379 if (!LI) 380 return nullptr; 381 Value *BaseAddr = LoadBaseAddresses[LI]; 382 if (!BaseAddr) 383 return nullptr; 384 385 auto GV = dyn_cast<GlobalVariable>(BaseAddr); 386 if (!GV) 387 return nullptr; 388 389 ConstantDataSequential *CDS = 390 dyn_cast<ConstantDataSequential>(GV->getInitializer()); 391 if (!CDS) 392 return nullptr; 393 394 const SCEV *BaseAddrSE = SE.getSCEV(BaseAddr); 395 const SCEV *S = SE.getSCEV(LI->getPointerOperand()); 396 const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE); 397 398 APInt StepC, StartC; 399 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE); 400 if (!AR) 401 return nullptr; 402 403 if (const SCEVConstant *StepSE = 404 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE))) 405 StepC = StepSE->getValue()->getValue(); 406 else 407 return nullptr; 408 409 if (const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart())) 410 StartC = StartSE->getValue()->getValue(); 411 else 412 return nullptr; 413 414 unsigned ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U; 415 unsigned Start = StartC.getLimitedValue(); 416 unsigned Step = StepC.getLimitedValue(); 417 418 unsigned Index = (Start + Step * Iteration) / ElemSize; 419 if (Index >= CDS->getNumElements()) 420 return nullptr; 421 422 Constant *CV = CDS->getElementAsConstant(Index); 423 424 return CV; 425 } 426 427 public: 428 UnrollAnalyzer(const Loop *L, unsigned TripCount, ScalarEvolution &SE, 429 const TargetTransformInfo &TTI) 430 : L(L), TripCount(TripCount), SE(SE), TTI(TTI), 431 NumberOfOptimizedInstructions(0) {} 432 433 // Visit all loads the loop L, and for those that, after complete loop 434 // unrolling, would have a constant address and it will point to a known 435 // constant initializer, record its base address for future use. It is used 436 // when we estimate number of potentially simplified instructions. 437 void findConstFoldableLoads() { 438 for (auto BB : L->getBlocks()) { 439 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 440 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 441 if (!LI->isSimple()) 442 continue; 443 Value *AddrOp = LI->getPointerOperand(); 444 const SCEV *S = SE.getSCEV(AddrOp); 445 FindConstantPointers Visitor(L, SE); 446 SCEVTraversal<FindConstantPointers> T(Visitor); 447 T.visitAll(S); 448 if (Visitor.IndexIsConstant && Visitor.LoadCanBeConstantFolded) { 449 LoadBaseAddresses[LI] = Visitor.BaseAddress; 450 } 451 } 452 } 453 } 454 } 455 456 // Given a list of loads that could be constant-folded (LoadBaseAddresses), 457 // estimate number of optimized instructions after substituting the concrete 458 // values for the given Iteration. Also track how many instructions become 459 // dead through this process. 460 unsigned estimateNumberOfOptimizedInstructions(unsigned Iteration) { 461 // We keep a set vector for the worklist so that we don't wast space in the 462 // worklist queuing up the same instruction repeatedly. This can happen due 463 // to multiple operands being the same instruction or due to the same 464 // instruction being an operand of lots of things that end up dead or 465 // simplified. 466 SmallSetVector<Instruction *, 8> Worklist; 467 468 // Clear the simplified values and counts for this iteration. 469 SimplifiedValues.clear(); 470 CountedInstructions.clear(); 471 NumberOfOptimizedInstructions = 0; 472 473 // We start by adding all loads to the worklist. 474 for (auto &LoadDescr : LoadBaseAddresses) { 475 LoadInst *LI = LoadDescr.first; 476 SimplifiedValues[LI] = computeLoadValue(LI, Iteration); 477 if (CountedInstructions.insert(LI).second) 478 NumberOfOptimizedInstructions += TTI.getUserCost(LI); 479 480 for (User *U : LI->users()) 481 Worklist.insert(cast<Instruction>(U)); 482 } 483 484 // And then we try to simplify every user of every instruction from the 485 // worklist. If we do simplify a user, add it to the worklist to process 486 // its users as well. 487 while (!Worklist.empty()) { 488 Instruction *I = Worklist.pop_back_val(); 489 if (!L->contains(I)) 490 continue; 491 if (!visit(I)) 492 continue; 493 for (User *U : I->users()) 494 Worklist.insert(cast<Instruction>(U)); 495 } 496 497 // Now that we know the potentially simplifed instructions, estimate number 498 // of instructions that would become dead if we do perform the 499 // simplification. 500 501 // The dead instructions are held in a separate set. This is used to 502 // prevent us from re-examining instructions and make sure we only count 503 // the benifit once. The worklist's internal set handles insertion 504 // deduplication. 505 SmallPtrSet<Instruction *, 16> DeadInstructions; 506 507 // Lambda to enque operands onto the worklist. 508 auto EnqueueOperands = [&](Instruction &I) { 509 for (auto *Op : I.operand_values()) 510 if (auto *OpI = dyn_cast<Instruction>(Op)) 511 if (!OpI->use_empty()) 512 Worklist.insert(OpI); 513 }; 514 515 // Start by initializing worklist with simplified instructions. 516 for (auto &FoldedKeyValue : SimplifiedValues) 517 if (auto *FoldedInst = dyn_cast<Instruction>(FoldedKeyValue.first)) { 518 DeadInstructions.insert(FoldedInst); 519 520 // Add each instruction operand of this dead instruction to the 521 // worklist. 522 EnqueueOperands(*FoldedInst); 523 } 524 525 // If a definition of an insn is only used by simplified or dead 526 // instructions, it's also dead. Check defs of all instructions from the 527 // worklist. 528 while (!Worklist.empty()) { 529 Instruction *I = Worklist.pop_back_val(); 530 if (!L->contains(I)) 531 continue; 532 if (DeadInstructions.count(I)) 533 continue; 534 535 if (std::all_of(I->user_begin(), I->user_end(), [&](User *U) { 536 return DeadInstructions.count(cast<Instruction>(U)); 537 })) { 538 NumberOfOptimizedInstructions += TTI.getUserCost(I); 539 DeadInstructions.insert(I); 540 EnqueueOperands(*I); 541 } 542 } 543 return NumberOfOptimizedInstructions; 544 } 545 }; 546 } // namespace 547 548 // Complete loop unrolling can make some loads constant, and we need to know if 549 // that would expose any further optimization opportunities. 550 // This routine estimates this optimization effect and returns the number of 551 // instructions, that potentially might be optimized away. 552 static unsigned 553 approximateNumberOfOptimizedInstructions(const Loop *L, ScalarEvolution &SE, 554 unsigned TripCount, 555 const TargetTransformInfo &TTI) { 556 if (!TripCount || !UnrollMaxIterationsCountToAnalyze) 557 return 0; 558 559 UnrollAnalyzer UA(L, TripCount, SE, TTI); 560 UA.findConstFoldableLoads(); 561 562 // Estimate number of instructions, that could be simplified if we replace a 563 // load with the corresponding constant. Since the same load will take 564 // different values on different iterations, we have to go through all loop's 565 // iterations here. To limit ourselves here, we check only first N 566 // iterations, and then scale the found number, if necessary. 567 unsigned IterationsNumberForEstimate = 568 std::min<unsigned>(UnrollMaxIterationsCountToAnalyze, TripCount); 569 unsigned NumberOfOptimizedInstructions = 0; 570 for (unsigned i = 0; i < IterationsNumberForEstimate; ++i) 571 NumberOfOptimizedInstructions += 572 UA.estimateNumberOfOptimizedInstructions(i); 573 574 NumberOfOptimizedInstructions *= TripCount / IterationsNumberForEstimate; 575 576 return NumberOfOptimizedInstructions; 577 } 578 579 /// ApproximateLoopSize - Approximate the size of the loop. 580 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, 581 bool &NotDuplicatable, 582 const TargetTransformInfo &TTI, 583 AssumptionCache *AC) { 584 SmallPtrSet<const Value *, 32> EphValues; 585 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 586 587 CodeMetrics Metrics; 588 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 589 I != E; ++I) 590 Metrics.analyzeBasicBlock(*I, TTI, EphValues); 591 NumCalls = Metrics.NumInlineCandidates; 592 NotDuplicatable = Metrics.notDuplicatable; 593 594 unsigned LoopSize = Metrics.NumInsts; 595 596 // Don't allow an estimate of size zero. This would allows unrolling of loops 597 // with huge iteration counts, which is a compile time problem even if it's 598 // not a problem for code quality. Also, the code using this size may assume 599 // that each loop has at least three instructions (likely a conditional 600 // branch, a comparison feeding that branch, and some kind of loop increment 601 // feeding that comparison instruction). 602 LoopSize = std::max(LoopSize, 3u); 603 604 return LoopSize; 605 } 606 607 // Returns the loop hint metadata node with the given name (for example, 608 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is 609 // returned. 610 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) { 611 if (MDNode *LoopID = L->getLoopID()) 612 return GetUnrollMetadata(LoopID, Name); 613 return nullptr; 614 } 615 616 // Returns true if the loop has an unroll(full) pragma. 617 static bool HasUnrollFullPragma(const Loop *L) { 618 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full"); 619 } 620 621 // Returns true if the loop has an unroll(disable) pragma. 622 static bool HasUnrollDisablePragma(const Loop *L) { 623 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable"); 624 } 625 626 // Returns true if the loop has an runtime unroll(disable) pragma. 627 static bool HasRuntimeUnrollDisablePragma(const Loop *L) { 628 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable"); 629 } 630 631 // If loop has an unroll_count pragma return the (necessarily 632 // positive) value from the pragma. Otherwise return 0. 633 static unsigned UnrollCountPragmaValue(const Loop *L) { 634 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count"); 635 if (MD) { 636 assert(MD->getNumOperands() == 2 && 637 "Unroll count hint metadata should have two operands."); 638 unsigned Count = 639 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 640 assert(Count >= 1 && "Unroll count must be positive."); 641 return Count; 642 } 643 return 0; 644 } 645 646 // Remove existing unroll metadata and add unroll disable metadata to 647 // indicate the loop has already been unrolled. This prevents a loop 648 // from being unrolled more than is directed by a pragma if the loop 649 // unrolling pass is run more than once (which it generally is). 650 static void SetLoopAlreadyUnrolled(Loop *L) { 651 MDNode *LoopID = L->getLoopID(); 652 if (!LoopID) return; 653 654 // First remove any existing loop unrolling metadata. 655 SmallVector<Metadata *, 4> MDs; 656 // Reserve first location for self reference to the LoopID metadata node. 657 MDs.push_back(nullptr); 658 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 659 bool IsUnrollMetadata = false; 660 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 661 if (MD) { 662 const MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 663 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll."); 664 } 665 if (!IsUnrollMetadata) 666 MDs.push_back(LoopID->getOperand(i)); 667 } 668 669 // Add unroll(disable) metadata to disable future unrolling. 670 LLVMContext &Context = L->getHeader()->getContext(); 671 SmallVector<Metadata *, 1> DisableOperands; 672 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable")); 673 MDNode *DisableNode = MDNode::get(Context, DisableOperands); 674 MDs.push_back(DisableNode); 675 676 MDNode *NewLoopID = MDNode::get(Context, MDs); 677 // Set operand 0 to refer to the loop id itself. 678 NewLoopID->replaceOperandWith(0, NewLoopID); 679 L->setLoopID(NewLoopID); 680 } 681 682 unsigned LoopUnroll::selectUnrollCount( 683 const Loop *L, unsigned TripCount, bool PragmaFullUnroll, 684 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP, 685 bool &SetExplicitly) { 686 SetExplicitly = true; 687 688 // User-specified count (either as a command-line option or 689 // constructor parameter) has highest precedence. 690 unsigned Count = UserCount ? CurrentCount : 0; 691 692 // If there is no user-specified count, unroll pragmas have the next 693 // highest precendence. 694 if (Count == 0) { 695 if (PragmaCount) { 696 Count = PragmaCount; 697 } else if (PragmaFullUnroll) { 698 Count = TripCount; 699 } 700 } 701 702 if (Count == 0) 703 Count = UP.Count; 704 705 if (Count == 0) { 706 SetExplicitly = false; 707 if (TripCount == 0) 708 // Runtime trip count. 709 Count = UnrollRuntimeCount; 710 else 711 // Conservative heuristic: if we know the trip count, see if we can 712 // completely unroll (subject to the threshold, checked below); otherwise 713 // try to find greatest modulo of the trip count which is still under 714 // threshold value. 715 Count = TripCount; 716 } 717 if (TripCount && Count > TripCount) 718 return TripCount; 719 return Count; 720 } 721 722 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 723 if (skipOptnoneFunction(L)) 724 return false; 725 726 Function &F = *L->getHeader()->getParent(); 727 728 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 729 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>(); 730 const TargetTransformInfo &TTI = 731 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 732 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 733 734 BasicBlock *Header = L->getHeader(); 735 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 736 << "] Loop %" << Header->getName() << "\n"); 737 738 if (HasUnrollDisablePragma(L)) { 739 return false; 740 } 741 bool PragmaFullUnroll = HasUnrollFullPragma(L); 742 unsigned PragmaCount = UnrollCountPragmaValue(L); 743 bool HasPragma = PragmaFullUnroll || PragmaCount > 0; 744 745 TargetTransformInfo::UnrollingPreferences UP; 746 getUnrollingPreferences(L, TTI, UP); 747 748 // Find trip count and trip multiple if count is not available 749 unsigned TripCount = 0; 750 unsigned TripMultiple = 1; 751 // If there are multiple exiting blocks but one of them is the latch, use the 752 // latch for the trip count estimation. Otherwise insist on a single exiting 753 // block for the trip count estimation. 754 BasicBlock *ExitingBlock = L->getLoopLatch(); 755 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock)) 756 ExitingBlock = L->getExitingBlock(); 757 if (ExitingBlock) { 758 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock); 759 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock); 760 } 761 762 // Select an initial unroll count. This may be reduced later based 763 // on size thresholds. 764 bool CountSetExplicitly; 765 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll, 766 PragmaCount, UP, CountSetExplicitly); 767 768 unsigned NumInlineCandidates; 769 bool notDuplicatable; 770 unsigned LoopSize = 771 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC); 772 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 773 774 // When computing the unrolled size, note that the conditional branch on the 775 // backedge and the comparison feeding it are not replicated like the rest of 776 // the loop body (which is why 2 is subtracted). 777 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2; 778 if (notDuplicatable) { 779 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable" 780 << " instructions.\n"); 781 return false; 782 } 783 if (NumInlineCandidates != 0) { 784 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 785 return false; 786 } 787 788 unsigned NumberOfOptimizedInstructions = 789 approximateNumberOfOptimizedInstructions(L, *SE, TripCount, TTI); 790 DEBUG(dbgs() << " Complete unrolling could save: " 791 << NumberOfOptimizedInstructions << "\n"); 792 793 unsigned Threshold, PartialThreshold; 794 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold, 795 NumberOfOptimizedInstructions); 796 797 // Given Count, TripCount and thresholds determine the type of 798 // unrolling which is to be performed. 799 enum { Full = 0, Partial = 1, Runtime = 2 }; 800 int Unrolling; 801 if (TripCount && Count == TripCount) { 802 if (Threshold != NoThreshold && UnrolledSize > Threshold) { 803 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count 804 << " because size: " << UnrolledSize << ">" << Threshold 805 << "\n"); 806 Unrolling = Partial; 807 } else { 808 Unrolling = Full; 809 } 810 } else if (TripCount && Count < TripCount) { 811 Unrolling = Partial; 812 } else { 813 Unrolling = Runtime; 814 } 815 816 // Reduce count based on the type of unrolling and the threshold values. 817 unsigned OriginalCount = Count; 818 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime; 819 if (HasRuntimeUnrollDisablePragma(L)) { 820 AllowRuntime = false; 821 } 822 if (Unrolling == Partial) { 823 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial; 824 if (!AllowPartial && !CountSetExplicitly) { 825 DEBUG(dbgs() << " will not try to unroll partially because " 826 << "-unroll-allow-partial not given\n"); 827 return false; 828 } 829 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) { 830 // Reduce unroll count to be modulo of TripCount for partial unrolling. 831 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2); 832 while (Count != 0 && TripCount % Count != 0) 833 Count--; 834 } 835 } else if (Unrolling == Runtime) { 836 if (!AllowRuntime && !CountSetExplicitly) { 837 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count " 838 << "-unroll-runtime not given\n"); 839 return false; 840 } 841 // Reduce unroll count to be the largest power-of-two factor of 842 // the original count which satisfies the threshold limit. 843 while (Count != 0 && UnrolledSize > PartialThreshold) { 844 Count >>= 1; 845 UnrolledSize = (LoopSize-2) * Count + 2; 846 } 847 if (Count > UP.MaxCount) 848 Count = UP.MaxCount; 849 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 850 } 851 852 if (HasPragma) { 853 if (PragmaCount != 0) 854 // If loop has an unroll count pragma mark loop as unrolled to prevent 855 // unrolling beyond that requested by the pragma. 856 SetLoopAlreadyUnrolled(L); 857 858 // Emit optimization remarks if we are unable to unroll the loop 859 // as directed by a pragma. 860 DebugLoc LoopLoc = L->getStartLoc(); 861 Function *F = Header->getParent(); 862 LLVMContext &Ctx = F->getContext(); 863 if (PragmaFullUnroll && PragmaCount == 0) { 864 if (TripCount && Count != TripCount) { 865 emitOptimizationRemarkMissed( 866 Ctx, DEBUG_TYPE, *F, LoopLoc, 867 "Unable to fully unroll loop as directed by unroll(full) pragma " 868 "because unrolled size is too large."); 869 } else if (!TripCount) { 870 emitOptimizationRemarkMissed( 871 Ctx, DEBUG_TYPE, *F, LoopLoc, 872 "Unable to fully unroll loop as directed by unroll(full) pragma " 873 "because loop has a runtime trip count."); 874 } 875 } else if (PragmaCount > 0 && Count != OriginalCount) { 876 emitOptimizationRemarkMissed( 877 Ctx, DEBUG_TYPE, *F, LoopLoc, 878 "Unable to unroll loop the number of times directed by " 879 "unroll_count pragma because unrolled size is too large."); 880 } 881 } 882 883 if (Unrolling != Full && Count < 2) { 884 // Partial unrolling by 1 is a nop. For full unrolling, a factor 885 // of 1 makes sense because loop control can be eliminated. 886 return false; 887 } 888 889 // Unroll the loop. 890 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount, 891 TripMultiple, LI, this, &LPM, &AC)) 892 return false; 893 894 return true; 895 } 896