1 //===- LoopUnroll.cpp - Loop unroller pass --------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This pass implements a simple loop unroller. It works best when loops have 10 // been canonicalized by the -indvars pass, allowing it to determine the trip 11 // counts of loops easily. 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Scalar/LoopUnrollPass.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/DenseMapInfo.h" 17 #include "llvm/ADT/DenseSet.h" 18 #include "llvm/ADT/None.h" 19 #include "llvm/ADT/Optional.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/ADT/SmallPtrSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/StringRef.h" 25 #include "llvm/Analysis/AssumptionCache.h" 26 #include "llvm/Analysis/BlockFrequencyInfo.h" 27 #include "llvm/Analysis/CodeMetrics.h" 28 #include "llvm/Analysis/LazyBlockFrequencyInfo.h" 29 #include "llvm/Analysis/LoopAnalysisManager.h" 30 #include "llvm/Analysis/LoopInfo.h" 31 #include "llvm/Analysis/LoopPass.h" 32 #include "llvm/Analysis/LoopUnrollAnalyzer.h" 33 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 34 #include "llvm/Analysis/ProfileSummaryInfo.h" 35 #include "llvm/Analysis/ScalarEvolution.h" 36 #include "llvm/Analysis/TargetTransformInfo.h" 37 #include "llvm/IR/BasicBlock.h" 38 #include "llvm/IR/CFG.h" 39 #include "llvm/IR/Constant.h" 40 #include "llvm/IR/Constants.h" 41 #include "llvm/IR/DiagnosticInfo.h" 42 #include "llvm/IR/Dominators.h" 43 #include "llvm/IR/Function.h" 44 #include "llvm/IR/Instruction.h" 45 #include "llvm/IR/Instructions.h" 46 #include "llvm/IR/IntrinsicInst.h" 47 #include "llvm/IR/Metadata.h" 48 #include "llvm/IR/PassManager.h" 49 #include "llvm/InitializePasses.h" 50 #include "llvm/Pass.h" 51 #include "llvm/Support/Casting.h" 52 #include "llvm/Support/CommandLine.h" 53 #include "llvm/Support/Debug.h" 54 #include "llvm/Support/ErrorHandling.h" 55 #include "llvm/Support/raw_ostream.h" 56 #include "llvm/Transforms/Scalar.h" 57 #include "llvm/Transforms/Scalar/LoopPassManager.h" 58 #include "llvm/Transforms/Utils.h" 59 #include "llvm/Transforms/Utils/LoopSimplify.h" 60 #include "llvm/Transforms/Utils/LoopUtils.h" 61 #include "llvm/Transforms/Utils/SizeOpts.h" 62 #include "llvm/Transforms/Utils/UnrollLoop.h" 63 #include <algorithm> 64 #include <cassert> 65 #include <cstdint> 66 #include <limits> 67 #include <string> 68 #include <tuple> 69 #include <utility> 70 71 using namespace llvm; 72 73 #define DEBUG_TYPE "loop-unroll" 74 75 cl::opt<bool> llvm::ForgetSCEVInLoopUnroll( 76 "forget-scev-loop-unroll", cl::init(false), cl::Hidden, 77 cl::desc("Forget everything in SCEV when doing LoopUnroll, instead of just" 78 " the current top-most loop. This is somtimes preferred to reduce" 79 " compile time.")); 80 81 static cl::opt<unsigned> 82 UnrollThreshold("unroll-threshold", cl::Hidden, 83 cl::desc("The cost threshold for loop unrolling")); 84 85 static cl::opt<unsigned> UnrollPartialThreshold( 86 "unroll-partial-threshold", cl::Hidden, 87 cl::desc("The cost threshold for partial loop unrolling")); 88 89 static cl::opt<unsigned> UnrollMaxPercentThresholdBoost( 90 "unroll-max-percent-threshold-boost", cl::init(400), cl::Hidden, 91 cl::desc("The maximum 'boost' (represented as a percentage >= 100) applied " 92 "to the threshold when aggressively unrolling a loop due to the " 93 "dynamic cost savings. If completely unrolling a loop will reduce " 94 "the total runtime from X to Y, we boost the loop unroll " 95 "threshold to DefaultThreshold*std::min(MaxPercentThresholdBoost, " 96 "X/Y). This limit avoids excessive code bloat.")); 97 98 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze( 99 "unroll-max-iteration-count-to-analyze", cl::init(10), cl::Hidden, 100 cl::desc("Don't allow loop unrolling to simulate more than this number of" 101 "iterations when checking full unroll profitability")); 102 103 static cl::opt<unsigned> UnrollCount( 104 "unroll-count", cl::Hidden, 105 cl::desc("Use this unroll count for all loops including those with " 106 "unroll_count pragma values, for testing purposes")); 107 108 static cl::opt<unsigned> UnrollMaxCount( 109 "unroll-max-count", cl::Hidden, 110 cl::desc("Set the max unroll count for partial and runtime unrolling, for" 111 "testing purposes")); 112 113 static cl::opt<unsigned> UnrollFullMaxCount( 114 "unroll-full-max-count", cl::Hidden, 115 cl::desc( 116 "Set the max unroll count for full unrolling, for testing purposes")); 117 118 static cl::opt<unsigned> UnrollPeelCount( 119 "unroll-peel-count", cl::Hidden, 120 cl::desc("Set the unroll peeling count, for testing purposes")); 121 122 static cl::opt<bool> 123 UnrollAllowPartial("unroll-allow-partial", cl::Hidden, 124 cl::desc("Allows loops to be partially unrolled until " 125 "-unroll-threshold loop size is reached.")); 126 127 static cl::opt<bool> UnrollAllowRemainder( 128 "unroll-allow-remainder", cl::Hidden, 129 cl::desc("Allow generation of a loop remainder (extra iterations) " 130 "when unrolling a loop.")); 131 132 static cl::opt<bool> 133 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::Hidden, 134 cl::desc("Unroll loops with run-time trip counts")); 135 136 static cl::opt<unsigned> UnrollMaxUpperBound( 137 "unroll-max-upperbound", cl::init(8), cl::Hidden, 138 cl::desc( 139 "The max of trip count upper bound that is considered in unrolling")); 140 141 static cl::opt<unsigned> PragmaUnrollThreshold( 142 "pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden, 143 cl::desc("Unrolled size limit for loops with an unroll(full) or " 144 "unroll_count pragma.")); 145 146 static cl::opt<unsigned> FlatLoopTripCountThreshold( 147 "flat-loop-tripcount-threshold", cl::init(5), cl::Hidden, 148 cl::desc("If the runtime tripcount for the loop is lower than the " 149 "threshold, the loop is considered as flat and will be less " 150 "aggressively unrolled.")); 151 152 static cl::opt<bool> 153 UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden, 154 cl::desc("Allows loops to be peeled when the dynamic " 155 "trip count is known to be low.")); 156 157 static cl::opt<bool> UnrollAllowLoopNestsPeeling( 158 "unroll-allow-loop-nests-peeling", cl::init(false), cl::Hidden, 159 cl::desc("Allows loop nests to be peeled.")); 160 161 static cl::opt<bool> UnrollUnrollRemainder( 162 "unroll-remainder", cl::Hidden, 163 cl::desc("Allow the loop remainder to be unrolled.")); 164 165 // This option isn't ever intended to be enabled, it serves to allow 166 // experiments to check the assumptions about when this kind of revisit is 167 // necessary. 168 static cl::opt<bool> UnrollRevisitChildLoops( 169 "unroll-revisit-child-loops", cl::Hidden, 170 cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. " 171 "This shouldn't typically be needed as child loops (or their " 172 "clones) were already visited.")); 173 174 static cl::opt<unsigned> UnrollThresholdAggressive( 175 "unroll-threshold-aggressive", cl::init(300), cl::Hidden, 176 cl::desc("Threshold (max size of unrolled loop) to use in aggressive (O3) " 177 "optimizations")); 178 static cl::opt<unsigned> 179 UnrollThresholdDefault("unroll-threshold-default", cl::init(150), 180 cl::Hidden, 181 cl::desc("Default threshold (max size of unrolled " 182 "loop), used in all but O3 optimizations")); 183 184 /// A magic value for use with the Threshold parameter to indicate 185 /// that the loop unroll should be performed regardless of how much 186 /// code expansion would result. 187 static const unsigned NoThreshold = std::numeric_limits<unsigned>::max(); 188 189 /// Gather the various unrolling parameters based on the defaults, compiler 190 /// flags, TTI overrides and user specified parameters. 191 TargetTransformInfo::UnrollingPreferences llvm::gatherUnrollingPreferences( 192 Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, 193 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, int OptLevel, 194 Optional<unsigned> UserThreshold, Optional<unsigned> UserCount, 195 Optional<bool> UserAllowPartial, Optional<bool> UserRuntime, 196 Optional<bool> UserUpperBound, Optional<unsigned> UserFullUnrollMaxCount) { 197 TargetTransformInfo::UnrollingPreferences UP; 198 199 // Set up the defaults 200 UP.Threshold = 201 OptLevel > 2 ? UnrollThresholdAggressive : UnrollThresholdDefault; 202 UP.MaxPercentThresholdBoost = 400; 203 UP.OptSizeThreshold = 0; 204 UP.PartialThreshold = 150; 205 UP.PartialOptSizeThreshold = 0; 206 UP.Count = 0; 207 UP.DefaultUnrollRuntimeCount = 8; 208 UP.MaxCount = std::numeric_limits<unsigned>::max(); 209 UP.FullUnrollMaxCount = std::numeric_limits<unsigned>::max(); 210 UP.BEInsns = 2; 211 UP.Partial = false; 212 UP.Runtime = false; 213 UP.AllowRemainder = true; 214 UP.UnrollRemainder = false; 215 UP.AllowExpensiveTripCount = false; 216 UP.Force = false; 217 UP.UpperBound = false; 218 UP.UnrollAndJam = false; 219 UP.UnrollAndJamInnerLoopThreshold = 60; 220 UP.MaxIterationsCountToAnalyze = UnrollMaxIterationsCountToAnalyze; 221 222 // Override with any target specific settings 223 TTI.getUnrollingPreferences(L, SE, UP); 224 225 // Apply size attributes 226 bool OptForSize = L->getHeader()->getParent()->hasOptSize() || 227 llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, 228 PGSOQueryType::IRPass); 229 if (OptForSize) { 230 UP.Threshold = UP.OptSizeThreshold; 231 UP.PartialThreshold = UP.PartialOptSizeThreshold; 232 UP.MaxPercentThresholdBoost = 100; 233 } 234 235 // Apply any user values specified by cl::opt 236 if (UnrollThreshold.getNumOccurrences() > 0) 237 UP.Threshold = UnrollThreshold; 238 if (UnrollPartialThreshold.getNumOccurrences() > 0) 239 UP.PartialThreshold = UnrollPartialThreshold; 240 if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0) 241 UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost; 242 if (UnrollMaxCount.getNumOccurrences() > 0) 243 UP.MaxCount = UnrollMaxCount; 244 if (UnrollFullMaxCount.getNumOccurrences() > 0) 245 UP.FullUnrollMaxCount = UnrollFullMaxCount; 246 if (UnrollAllowPartial.getNumOccurrences() > 0) 247 UP.Partial = UnrollAllowPartial; 248 if (UnrollAllowRemainder.getNumOccurrences() > 0) 249 UP.AllowRemainder = UnrollAllowRemainder; 250 if (UnrollRuntime.getNumOccurrences() > 0) 251 UP.Runtime = UnrollRuntime; 252 if (UnrollMaxUpperBound == 0) 253 UP.UpperBound = false; 254 if (UnrollUnrollRemainder.getNumOccurrences() > 0) 255 UP.UnrollRemainder = UnrollUnrollRemainder; 256 if (UnrollMaxIterationsCountToAnalyze.getNumOccurrences() > 0) 257 UP.MaxIterationsCountToAnalyze = UnrollMaxIterationsCountToAnalyze; 258 259 // Apply user values provided by argument 260 if (UserThreshold.hasValue()) { 261 UP.Threshold = *UserThreshold; 262 UP.PartialThreshold = *UserThreshold; 263 } 264 if (UserCount.hasValue()) 265 UP.Count = *UserCount; 266 if (UserAllowPartial.hasValue()) 267 UP.Partial = *UserAllowPartial; 268 if (UserRuntime.hasValue()) 269 UP.Runtime = *UserRuntime; 270 if (UserUpperBound.hasValue()) 271 UP.UpperBound = *UserUpperBound; 272 if (UserFullUnrollMaxCount.hasValue()) 273 UP.FullUnrollMaxCount = *UserFullUnrollMaxCount; 274 275 return UP; 276 } 277 278 TargetTransformInfo::PeelingPreferences 279 llvm::gatherPeelingPreferences(Loop *L, ScalarEvolution &SE, 280 const TargetTransformInfo &TTI, 281 Optional<bool> UserAllowPeeling, 282 Optional<bool> UserAllowProfileBasedPeeling) { 283 TargetTransformInfo::PeelingPreferences PP; 284 285 // Default values 286 PP.PeelCount = 0; 287 PP.AllowPeeling = true; 288 PP.AllowLoopNestsPeeling = false; 289 PP.PeelProfiledIterations = true; 290 291 // Get Target Specifc Values 292 TTI.getPeelingPreferences(L, SE, PP); 293 294 // User Specified Values using cl::opt 295 if (UnrollPeelCount.getNumOccurrences() > 0) 296 PP.PeelCount = UnrollPeelCount; 297 if (UnrollAllowPeeling.getNumOccurrences() > 0) 298 PP.AllowPeeling = UnrollAllowPeeling; 299 if (UnrollAllowLoopNestsPeeling.getNumOccurrences() > 0) 300 PP.AllowLoopNestsPeeling = UnrollAllowLoopNestsPeeling; 301 302 // User Specifed values provided by argument 303 if (UserAllowPeeling.hasValue()) 304 PP.AllowPeeling = *UserAllowPeeling; 305 if (UserAllowProfileBasedPeeling.hasValue()) 306 PP.PeelProfiledIterations = *UserAllowProfileBasedPeeling; 307 308 return PP; 309 } 310 311 namespace { 312 313 /// A struct to densely store the state of an instruction after unrolling at 314 /// each iteration. 315 /// 316 /// This is designed to work like a tuple of <Instruction *, int> for the 317 /// purposes of hashing and lookup, but to be able to associate two boolean 318 /// states with each key. 319 struct UnrolledInstState { 320 Instruction *I; 321 int Iteration : 30; 322 unsigned IsFree : 1; 323 unsigned IsCounted : 1; 324 }; 325 326 /// Hashing and equality testing for a set of the instruction states. 327 struct UnrolledInstStateKeyInfo { 328 using PtrInfo = DenseMapInfo<Instruction *>; 329 using PairInfo = DenseMapInfo<std::pair<Instruction *, int>>; 330 331 static inline UnrolledInstState getEmptyKey() { 332 return {PtrInfo::getEmptyKey(), 0, 0, 0}; 333 } 334 335 static inline UnrolledInstState getTombstoneKey() { 336 return {PtrInfo::getTombstoneKey(), 0, 0, 0}; 337 } 338 339 static inline unsigned getHashValue(const UnrolledInstState &S) { 340 return PairInfo::getHashValue({S.I, S.Iteration}); 341 } 342 343 static inline bool isEqual(const UnrolledInstState &LHS, 344 const UnrolledInstState &RHS) { 345 return PairInfo::isEqual({LHS.I, LHS.Iteration}, {RHS.I, RHS.Iteration}); 346 } 347 }; 348 349 struct EstimatedUnrollCost { 350 /// The estimated cost after unrolling. 351 unsigned UnrolledCost; 352 353 /// The estimated dynamic cost of executing the instructions in the 354 /// rolled form. 355 unsigned RolledDynamicCost; 356 }; 357 358 } // end anonymous namespace 359 360 /// Figure out if the loop is worth full unrolling. 361 /// 362 /// Complete loop unrolling can make some loads constant, and we need to know 363 /// if that would expose any further optimization opportunities. This routine 364 /// estimates this optimization. It computes cost of unrolled loop 365 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By 366 /// dynamic cost we mean that we won't count costs of blocks that are known not 367 /// to be executed (i.e. if we have a branch in the loop and we know that at the 368 /// given iteration its condition would be resolved to true, we won't add up the 369 /// cost of the 'false'-block). 370 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If 371 /// the analysis failed (no benefits expected from the unrolling, or the loop is 372 /// too big to analyze), the returned value is None. 373 static Optional<EstimatedUnrollCost> analyzeLoopUnrollCost( 374 const Loop *L, unsigned TripCount, DominatorTree &DT, ScalarEvolution &SE, 375 const SmallPtrSetImpl<const Value *> &EphValues, 376 const TargetTransformInfo &TTI, unsigned MaxUnrolledLoopSize, 377 unsigned MaxIterationsCountToAnalyze) { 378 // We want to be able to scale offsets by the trip count and add more offsets 379 // to them without checking for overflows, and we already don't want to 380 // analyze *massive* trip counts, so we force the max to be reasonably small. 381 assert(MaxIterationsCountToAnalyze < 382 (unsigned)(std::numeric_limits<int>::max() / 2) && 383 "The unroll iterations max is too large!"); 384 385 // Only analyze inner loops. We can't properly estimate cost of nested loops 386 // and we won't visit inner loops again anyway. 387 if (!L->empty()) 388 return None; 389 390 // Don't simulate loops with a big or unknown tripcount 391 if (!TripCount || TripCount > MaxIterationsCountToAnalyze) 392 return None; 393 394 SmallSetVector<BasicBlock *, 16> BBWorklist; 395 SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist; 396 DenseMap<Value *, Constant *> SimplifiedValues; 397 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues; 398 399 // The estimated cost of the unrolled form of the loop. We try to estimate 400 // this by simplifying as much as we can while computing the estimate. 401 unsigned UnrolledCost = 0; 402 403 // We also track the estimated dynamic (that is, actually executed) cost in 404 // the rolled form. This helps identify cases when the savings from unrolling 405 // aren't just exposing dead control flows, but actual reduced dynamic 406 // instructions due to the simplifications which we expect to occur after 407 // unrolling. 408 unsigned RolledDynamicCost = 0; 409 410 // We track the simplification of each instruction in each iteration. We use 411 // this to recursively merge costs into the unrolled cost on-demand so that 412 // we don't count the cost of any dead code. This is essentially a map from 413 // <instruction, int> to <bool, bool>, but stored as a densely packed struct. 414 DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap; 415 416 // A small worklist used to accumulate cost of instructions from each 417 // observable and reached root in the loop. 418 SmallVector<Instruction *, 16> CostWorklist; 419 420 // PHI-used worklist used between iterations while accumulating cost. 421 SmallVector<Instruction *, 4> PHIUsedList; 422 423 // Helper function to accumulate cost for instructions in the loop. 424 auto AddCostRecursively = [&](Instruction &RootI, int Iteration) { 425 assert(Iteration >= 0 && "Cannot have a negative iteration!"); 426 assert(CostWorklist.empty() && "Must start with an empty cost list"); 427 assert(PHIUsedList.empty() && "Must start with an empty phi used list"); 428 CostWorklist.push_back(&RootI); 429 for (;; --Iteration) { 430 do { 431 Instruction *I = CostWorklist.pop_back_val(); 432 433 // InstCostMap only uses I and Iteration as a key, the other two values 434 // don't matter here. 435 auto CostIter = InstCostMap.find({I, Iteration, 0, 0}); 436 if (CostIter == InstCostMap.end()) 437 // If an input to a PHI node comes from a dead path through the loop 438 // we may have no cost data for it here. What that actually means is 439 // that it is free. 440 continue; 441 auto &Cost = *CostIter; 442 if (Cost.IsCounted) 443 // Already counted this instruction. 444 continue; 445 446 // Mark that we are counting the cost of this instruction now. 447 Cost.IsCounted = true; 448 449 // If this is a PHI node in the loop header, just add it to the PHI set. 450 if (auto *PhiI = dyn_cast<PHINode>(I)) 451 if (PhiI->getParent() == L->getHeader()) { 452 assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they " 453 "inherently simplify during unrolling."); 454 if (Iteration == 0) 455 continue; 456 457 // Push the incoming value from the backedge into the PHI used list 458 // if it is an in-loop instruction. We'll use this to populate the 459 // cost worklist for the next iteration (as we count backwards). 460 if (auto *OpI = dyn_cast<Instruction>( 461 PhiI->getIncomingValueForBlock(L->getLoopLatch()))) 462 if (L->contains(OpI)) 463 PHIUsedList.push_back(OpI); 464 continue; 465 } 466 467 // First accumulate the cost of this instruction. 468 if (!Cost.IsFree) { 469 UnrolledCost += TTI.getUserCost(I, TargetTransformInfo::TCK_CodeSize); 470 LLVM_DEBUG(dbgs() << "Adding cost of instruction (iteration " 471 << Iteration << "): "); 472 LLVM_DEBUG(I->dump()); 473 } 474 475 // We must count the cost of every operand which is not free, 476 // recursively. If we reach a loop PHI node, simply add it to the set 477 // to be considered on the next iteration (backwards!). 478 for (Value *Op : I->operands()) { 479 // Check whether this operand is free due to being a constant or 480 // outside the loop. 481 auto *OpI = dyn_cast<Instruction>(Op); 482 if (!OpI || !L->contains(OpI)) 483 continue; 484 485 // Otherwise accumulate its cost. 486 CostWorklist.push_back(OpI); 487 } 488 } while (!CostWorklist.empty()); 489 490 if (PHIUsedList.empty()) 491 // We've exhausted the search. 492 break; 493 494 assert(Iteration > 0 && 495 "Cannot track PHI-used values past the first iteration!"); 496 CostWorklist.append(PHIUsedList.begin(), PHIUsedList.end()); 497 PHIUsedList.clear(); 498 } 499 }; 500 501 // Ensure that we don't violate the loop structure invariants relied on by 502 // this analysis. 503 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first."); 504 assert(L->isLCSSAForm(DT) && 505 "Must have loops in LCSSA form to track live-out values."); 506 507 LLVM_DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n"); 508 509 // Simulate execution of each iteration of the loop counting instructions, 510 // which would be simplified. 511 // Since the same load will take different values on different iterations, 512 // we literally have to go through all loop's iterations. 513 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) { 514 LLVM_DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n"); 515 516 // Prepare for the iteration by collecting any simplified entry or backedge 517 // inputs. 518 for (Instruction &I : *L->getHeader()) { 519 auto *PHI = dyn_cast<PHINode>(&I); 520 if (!PHI) 521 break; 522 523 // The loop header PHI nodes must have exactly two input: one from the 524 // loop preheader and one from the loop latch. 525 assert( 526 PHI->getNumIncomingValues() == 2 && 527 "Must have an incoming value only for the preheader and the latch."); 528 529 Value *V = PHI->getIncomingValueForBlock( 530 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch()); 531 Constant *C = dyn_cast<Constant>(V); 532 if (Iteration != 0 && !C) 533 C = SimplifiedValues.lookup(V); 534 if (C) 535 SimplifiedInputValues.push_back({PHI, C}); 536 } 537 538 // Now clear and re-populate the map for the next iteration. 539 SimplifiedValues.clear(); 540 while (!SimplifiedInputValues.empty()) 541 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val()); 542 543 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L); 544 545 BBWorklist.clear(); 546 BBWorklist.insert(L->getHeader()); 547 // Note that we *must not* cache the size, this loop grows the worklist. 548 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) { 549 BasicBlock *BB = BBWorklist[Idx]; 550 551 // Visit all instructions in the given basic block and try to simplify 552 // it. We don't change the actual IR, just count optimization 553 // opportunities. 554 for (Instruction &I : *BB) { 555 // These won't get into the final code - don't even try calculating the 556 // cost for them. 557 if (isa<DbgInfoIntrinsic>(I) || EphValues.count(&I)) 558 continue; 559 560 // Track this instruction's expected baseline cost when executing the 561 // rolled loop form. 562 RolledDynamicCost += TTI.getUserCost(&I, TargetTransformInfo::TCK_CodeSize); 563 564 // Visit the instruction to analyze its loop cost after unrolling, 565 // and if the visitor returns true, mark the instruction as free after 566 // unrolling and continue. 567 bool IsFree = Analyzer.visit(I); 568 bool Inserted = InstCostMap.insert({&I, (int)Iteration, 569 (unsigned)IsFree, 570 /*IsCounted*/ false}).second; 571 (void)Inserted; 572 assert(Inserted && "Cannot have a state for an unvisited instruction!"); 573 574 if (IsFree) 575 continue; 576 577 // Can't properly model a cost of a call. 578 // FIXME: With a proper cost model we should be able to do it. 579 if (auto *CI = dyn_cast<CallInst>(&I)) { 580 const Function *Callee = CI->getCalledFunction(); 581 if (!Callee || TTI.isLoweredToCall(Callee)) { 582 LLVM_DEBUG(dbgs() << "Can't analyze cost of loop with call\n"); 583 return None; 584 } 585 } 586 587 // If the instruction might have a side-effect recursively account for 588 // the cost of it and all the instructions leading up to it. 589 if (I.mayHaveSideEffects()) 590 AddCostRecursively(I, Iteration); 591 592 // If unrolled body turns out to be too big, bail out. 593 if (UnrolledCost > MaxUnrolledLoopSize) { 594 LLVM_DEBUG(dbgs() << " Exceeded threshold.. exiting.\n" 595 << " UnrolledCost: " << UnrolledCost 596 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize 597 << "\n"); 598 return None; 599 } 600 } 601 602 Instruction *TI = BB->getTerminator(); 603 604 // Add in the live successors by first checking whether we have terminator 605 // that may be simplified based on the values simplified by this call. 606 BasicBlock *KnownSucc = nullptr; 607 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 608 if (BI->isConditional()) { 609 if (Constant *SimpleCond = 610 SimplifiedValues.lookup(BI->getCondition())) { 611 // Just take the first successor if condition is undef 612 if (isa<UndefValue>(SimpleCond)) 613 KnownSucc = BI->getSuccessor(0); 614 else if (ConstantInt *SimpleCondVal = 615 dyn_cast<ConstantInt>(SimpleCond)) 616 KnownSucc = BI->getSuccessor(SimpleCondVal->isZero() ? 1 : 0); 617 } 618 } 619 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 620 if (Constant *SimpleCond = 621 SimplifiedValues.lookup(SI->getCondition())) { 622 // Just take the first successor if condition is undef 623 if (isa<UndefValue>(SimpleCond)) 624 KnownSucc = SI->getSuccessor(0); 625 else if (ConstantInt *SimpleCondVal = 626 dyn_cast<ConstantInt>(SimpleCond)) 627 KnownSucc = SI->findCaseValue(SimpleCondVal)->getCaseSuccessor(); 628 } 629 } 630 if (KnownSucc) { 631 if (L->contains(KnownSucc)) 632 BBWorklist.insert(KnownSucc); 633 else 634 ExitWorklist.insert({BB, KnownSucc}); 635 continue; 636 } 637 638 // Add BB's successors to the worklist. 639 for (BasicBlock *Succ : successors(BB)) 640 if (L->contains(Succ)) 641 BBWorklist.insert(Succ); 642 else 643 ExitWorklist.insert({BB, Succ}); 644 AddCostRecursively(*TI, Iteration); 645 } 646 647 // If we found no optimization opportunities on the first iteration, we 648 // won't find them on later ones too. 649 if (UnrolledCost == RolledDynamicCost) { 650 LLVM_DEBUG(dbgs() << " No opportunities found.. exiting.\n" 651 << " UnrolledCost: " << UnrolledCost << "\n"); 652 return None; 653 } 654 } 655 656 while (!ExitWorklist.empty()) { 657 BasicBlock *ExitingBB, *ExitBB; 658 std::tie(ExitingBB, ExitBB) = ExitWorklist.pop_back_val(); 659 660 for (Instruction &I : *ExitBB) { 661 auto *PN = dyn_cast<PHINode>(&I); 662 if (!PN) 663 break; 664 665 Value *Op = PN->getIncomingValueForBlock(ExitingBB); 666 if (auto *OpI = dyn_cast<Instruction>(Op)) 667 if (L->contains(OpI)) 668 AddCostRecursively(*OpI, TripCount - 1); 669 } 670 } 671 672 LLVM_DEBUG(dbgs() << "Analysis finished:\n" 673 << "UnrolledCost: " << UnrolledCost << ", " 674 << "RolledDynamicCost: " << RolledDynamicCost << "\n"); 675 return {{UnrolledCost, RolledDynamicCost}}; 676 } 677 678 /// ApproximateLoopSize - Approximate the size of the loop. 679 unsigned llvm::ApproximateLoopSize( 680 const Loop *L, unsigned &NumCalls, bool &NotDuplicatable, bool &Convergent, 681 const TargetTransformInfo &TTI, 682 const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) { 683 CodeMetrics Metrics; 684 for (BasicBlock *BB : L->blocks()) 685 Metrics.analyzeBasicBlock(BB, TTI, EphValues); 686 NumCalls = Metrics.NumInlineCandidates; 687 NotDuplicatable = Metrics.notDuplicatable; 688 Convergent = Metrics.convergent; 689 690 unsigned LoopSize = Metrics.NumInsts; 691 692 // Don't allow an estimate of size zero. This would allows unrolling of loops 693 // with huge iteration counts, which is a compile time problem even if it's 694 // not a problem for code quality. Also, the code using this size may assume 695 // that each loop has at least three instructions (likely a conditional 696 // branch, a comparison feeding that branch, and some kind of loop increment 697 // feeding that comparison instruction). 698 LoopSize = std::max(LoopSize, BEInsns + 1); 699 700 return LoopSize; 701 } 702 703 // Returns the loop hint metadata node with the given name (for example, 704 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is 705 // returned. 706 static MDNode *getUnrollMetadataForLoop(const Loop *L, StringRef Name) { 707 if (MDNode *LoopID = L->getLoopID()) 708 return GetUnrollMetadata(LoopID, Name); 709 return nullptr; 710 } 711 712 // Returns true if the loop has an unroll(full) pragma. 713 static bool hasUnrollFullPragma(const Loop *L) { 714 return getUnrollMetadataForLoop(L, "llvm.loop.unroll.full"); 715 } 716 717 // Returns true if the loop has an unroll(enable) pragma. This metadata is used 718 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives. 719 static bool hasUnrollEnablePragma(const Loop *L) { 720 return getUnrollMetadataForLoop(L, "llvm.loop.unroll.enable"); 721 } 722 723 // Returns true if the loop has an runtime unroll(disable) pragma. 724 static bool hasRuntimeUnrollDisablePragma(const Loop *L) { 725 return getUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable"); 726 } 727 728 // If loop has an unroll_count pragma return the (necessarily 729 // positive) value from the pragma. Otherwise return 0. 730 static unsigned unrollCountPragmaValue(const Loop *L) { 731 MDNode *MD = getUnrollMetadataForLoop(L, "llvm.loop.unroll.count"); 732 if (MD) { 733 assert(MD->getNumOperands() == 2 && 734 "Unroll count hint metadata should have two operands."); 735 unsigned Count = 736 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 737 assert(Count >= 1 && "Unroll count must be positive."); 738 return Count; 739 } 740 return 0; 741 } 742 743 // Computes the boosting factor for complete unrolling. 744 // If fully unrolling the loop would save a lot of RolledDynamicCost, it would 745 // be beneficial to fully unroll the loop even if unrolledcost is large. We 746 // use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust 747 // the unroll threshold. 748 static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost, 749 unsigned MaxPercentThresholdBoost) { 750 if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100) 751 return 100; 752 else if (Cost.UnrolledCost != 0) 753 // The boosting factor is RolledDynamicCost / UnrolledCost 754 return std::min(100 * Cost.RolledDynamicCost / Cost.UnrolledCost, 755 MaxPercentThresholdBoost); 756 else 757 return MaxPercentThresholdBoost; 758 } 759 760 // Returns loop size estimation for unrolled loop. 761 static uint64_t getUnrolledLoopSize( 762 unsigned LoopSize, 763 TargetTransformInfo::UnrollingPreferences &UP) { 764 assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!"); 765 return (uint64_t)(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns; 766 } 767 768 // Returns true if unroll count was set explicitly. 769 // Calculates unroll count and writes it to UP.Count. 770 // Unless IgnoreUser is true, will also use metadata and command-line options 771 // that are specific to to the LoopUnroll pass (which, for instance, are 772 // irrelevant for the LoopUnrollAndJam pass). 773 // FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes 774 // many LoopUnroll-specific options. The shared functionality should be 775 // refactored into it own function. 776 bool llvm::computeUnrollCount( 777 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI, 778 ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues, 779 OptimizationRemarkEmitter *ORE, unsigned &TripCount, unsigned MaxTripCount, 780 bool MaxOrZero, unsigned &TripMultiple, unsigned LoopSize, 781 TargetTransformInfo::UnrollingPreferences &UP, 782 TargetTransformInfo::PeelingPreferences &PP, bool &UseUpperBound) { 783 784 // Check for explicit Count. 785 // 1st priority is unroll count set by "unroll-count" option. 786 bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0; 787 if (UserUnrollCount) { 788 UP.Count = UnrollCount; 789 UP.AllowExpensiveTripCount = true; 790 UP.Force = true; 791 if (UP.AllowRemainder && getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) 792 return true; 793 } 794 795 // 2nd priority is unroll count set by pragma. 796 unsigned PragmaCount = unrollCountPragmaValue(L); 797 if (PragmaCount > 0) { 798 UP.Count = PragmaCount; 799 UP.Runtime = true; 800 UP.AllowExpensiveTripCount = true; 801 UP.Force = true; 802 if ((UP.AllowRemainder || (TripMultiple % PragmaCount == 0)) && 803 getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold) 804 return true; 805 } 806 bool PragmaFullUnroll = hasUnrollFullPragma(L); 807 if (PragmaFullUnroll && TripCount != 0) { 808 UP.Count = TripCount; 809 if (getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold) 810 return false; 811 } 812 813 bool PragmaEnableUnroll = hasUnrollEnablePragma(L); 814 bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll || 815 PragmaEnableUnroll || UserUnrollCount; 816 817 if (ExplicitUnroll && TripCount != 0) { 818 // If the loop has an unrolling pragma, we want to be more aggressive with 819 // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold 820 // value which is larger than the default limits. 821 UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold); 822 UP.PartialThreshold = 823 std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold); 824 } 825 826 // 3rd priority is full unroll count. 827 // Full unroll makes sense only when TripCount or its upper bound could be 828 // statically calculated. 829 // Also we need to check if we exceed FullUnrollMaxCount. 830 // If using the upper bound to unroll, TripMultiple should be set to 1 because 831 // we do not know when loop may exit. 832 833 // We can unroll by the upper bound amount if it's generally allowed or if 834 // we know that the loop is executed either the upper bound or zero times. 835 // (MaxOrZero unrolling keeps only the first loop test, so the number of 836 // loop tests remains the same compared to the non-unrolled version, whereas 837 // the generic upper bound unrolling keeps all but the last loop test so the 838 // number of loop tests goes up which may end up being worse on targets with 839 // constrained branch predictor resources so is controlled by an option.) 840 // In addition we only unroll small upper bounds. 841 unsigned FullUnrollMaxTripCount = MaxTripCount; 842 if (!(UP.UpperBound || MaxOrZero) || 843 FullUnrollMaxTripCount > UnrollMaxUpperBound) 844 FullUnrollMaxTripCount = 0; 845 846 // UnrollByMaxCount and ExactTripCount cannot both be non zero since we only 847 // compute the former when the latter is zero. 848 unsigned ExactTripCount = TripCount; 849 assert((ExactTripCount == 0 || FullUnrollMaxTripCount == 0) && 850 "ExtractTripCount and UnrollByMaxCount cannot both be non zero."); 851 852 unsigned FullUnrollTripCount = 853 ExactTripCount ? ExactTripCount : FullUnrollMaxTripCount; 854 UP.Count = FullUnrollTripCount; 855 if (FullUnrollTripCount && FullUnrollTripCount <= UP.FullUnrollMaxCount) { 856 // When computing the unrolled size, note that BEInsns are not replicated 857 // like the rest of the loop body. 858 if (getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) { 859 UseUpperBound = (FullUnrollMaxTripCount == FullUnrollTripCount); 860 TripCount = FullUnrollTripCount; 861 TripMultiple = UP.UpperBound ? 1 : TripMultiple; 862 return ExplicitUnroll; 863 } else { 864 // The loop isn't that small, but we still can fully unroll it if that 865 // helps to remove a significant number of instructions. 866 // To check that, run additional analysis on the loop. 867 if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost( 868 L, FullUnrollTripCount, DT, SE, EphValues, TTI, 869 UP.Threshold * UP.MaxPercentThresholdBoost / 100, 870 UP.MaxIterationsCountToAnalyze)) { 871 unsigned Boost = 872 getFullUnrollBoostingFactor(*Cost, UP.MaxPercentThresholdBoost); 873 if (Cost->UnrolledCost < UP.Threshold * Boost / 100) { 874 UseUpperBound = (FullUnrollMaxTripCount == FullUnrollTripCount); 875 TripCount = FullUnrollTripCount; 876 TripMultiple = UP.UpperBound ? 1 : TripMultiple; 877 return ExplicitUnroll; 878 } 879 } 880 } 881 } 882 883 // 4th priority is loop peeling. 884 computePeelCount(L, LoopSize, UP, PP, TripCount, SE); 885 if (PP.PeelCount) { 886 UP.Runtime = false; 887 UP.Count = 1; 888 return ExplicitUnroll; 889 } 890 891 // 5th priority is partial unrolling. 892 // Try partial unroll only when TripCount could be statically calculated. 893 if (TripCount) { 894 UP.Partial |= ExplicitUnroll; 895 if (!UP.Partial) { 896 LLVM_DEBUG(dbgs() << " will not try to unroll partially because " 897 << "-unroll-allow-partial not given\n"); 898 UP.Count = 0; 899 return false; 900 } 901 if (UP.Count == 0) 902 UP.Count = TripCount; 903 if (UP.PartialThreshold != NoThreshold) { 904 // Reduce unroll count to be modulo of TripCount for partial unrolling. 905 if (getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold) 906 UP.Count = 907 (std::max(UP.PartialThreshold, UP.BEInsns + 1) - UP.BEInsns) / 908 (LoopSize - UP.BEInsns); 909 if (UP.Count > UP.MaxCount) 910 UP.Count = UP.MaxCount; 911 while (UP.Count != 0 && TripCount % UP.Count != 0) 912 UP.Count--; 913 if (UP.AllowRemainder && UP.Count <= 1) { 914 // If there is no Count that is modulo of TripCount, set Count to 915 // largest power-of-two factor that satisfies the threshold limit. 916 // As we'll create fixup loop, do the type of unrolling only if 917 // remainder loop is allowed. 918 UP.Count = UP.DefaultUnrollRuntimeCount; 919 while (UP.Count != 0 && 920 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold) 921 UP.Count >>= 1; 922 } 923 if (UP.Count < 2) { 924 if (PragmaEnableUnroll) 925 ORE->emit([&]() { 926 return OptimizationRemarkMissed(DEBUG_TYPE, 927 "UnrollAsDirectedTooLarge", 928 L->getStartLoc(), L->getHeader()) 929 << "Unable to unroll loop as directed by unroll(enable) " 930 "pragma " 931 "because unrolled size is too large."; 932 }); 933 UP.Count = 0; 934 } 935 } else { 936 UP.Count = TripCount; 937 } 938 if (UP.Count > UP.MaxCount) 939 UP.Count = UP.MaxCount; 940 if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount && 941 UP.Count != TripCount) 942 ORE->emit([&]() { 943 return OptimizationRemarkMissed(DEBUG_TYPE, 944 "FullUnrollAsDirectedTooLarge", 945 L->getStartLoc(), L->getHeader()) 946 << "Unable to fully unroll loop as directed by unroll pragma " 947 "because " 948 "unrolled size is too large."; 949 }); 950 LLVM_DEBUG(dbgs() << " partially unrolling with count: " << UP.Count 951 << "\n"); 952 return ExplicitUnroll; 953 } 954 assert(TripCount == 0 && 955 "All cases when TripCount is constant should be covered here."); 956 if (PragmaFullUnroll) 957 ORE->emit([&]() { 958 return OptimizationRemarkMissed( 959 DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount", 960 L->getStartLoc(), L->getHeader()) 961 << "Unable to fully unroll loop as directed by unroll(full) " 962 "pragma " 963 "because loop has a runtime trip count."; 964 }); 965 966 // 6th priority is runtime unrolling. 967 // Don't unroll a runtime trip count loop when it is disabled. 968 if (hasRuntimeUnrollDisablePragma(L)) { 969 UP.Count = 0; 970 return false; 971 } 972 973 // Don't unroll a small upper bound loop unless user or TTI asked to do so. 974 if (MaxTripCount && !UP.Force && MaxTripCount < UnrollMaxUpperBound) { 975 UP.Count = 0; 976 return false; 977 } 978 979 // Check if the runtime trip count is too small when profile is available. 980 if (L->getHeader()->getParent()->hasProfileData()) { 981 if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) { 982 if (*ProfileTripCount < FlatLoopTripCountThreshold) 983 return false; 984 else 985 UP.AllowExpensiveTripCount = true; 986 } 987 } 988 989 // Reduce count based on the type of unrolling and the threshold values. 990 UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount; 991 if (!UP.Runtime) { 992 LLVM_DEBUG( 993 dbgs() << " will not try to unroll loop with runtime trip count " 994 << "-unroll-runtime not given\n"); 995 UP.Count = 0; 996 return false; 997 } 998 if (UP.Count == 0) 999 UP.Count = UP.DefaultUnrollRuntimeCount; 1000 1001 // Reduce unroll count to be the largest power-of-two factor of 1002 // the original count which satisfies the threshold limit. 1003 while (UP.Count != 0 && 1004 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold) 1005 UP.Count >>= 1; 1006 1007 #ifndef NDEBUG 1008 unsigned OrigCount = UP.Count; 1009 #endif 1010 1011 if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) { 1012 while (UP.Count != 0 && TripMultiple % UP.Count != 0) 1013 UP.Count >>= 1; 1014 LLVM_DEBUG( 1015 dbgs() << "Remainder loop is restricted (that could architecture " 1016 "specific or because the loop contains a convergent " 1017 "instruction), so unroll count must divide the trip " 1018 "multiple, " 1019 << TripMultiple << ". Reducing unroll count from " << OrigCount 1020 << " to " << UP.Count << ".\n"); 1021 1022 using namespace ore; 1023 1024 if (PragmaCount > 0 && !UP.AllowRemainder) 1025 ORE->emit([&]() { 1026 return OptimizationRemarkMissed(DEBUG_TYPE, 1027 "DifferentUnrollCountFromDirected", 1028 L->getStartLoc(), L->getHeader()) 1029 << "Unable to unroll loop the number of times directed by " 1030 "unroll_count pragma because remainder loop is restricted " 1031 "(that could architecture specific or because the loop " 1032 "contains a convergent instruction) and so must have an " 1033 "unroll " 1034 "count that divides the loop trip multiple of " 1035 << NV("TripMultiple", TripMultiple) << ". Unrolling instead " 1036 << NV("UnrollCount", UP.Count) << " time(s)."; 1037 }); 1038 } 1039 1040 if (UP.Count > UP.MaxCount) 1041 UP.Count = UP.MaxCount; 1042 1043 if (MaxTripCount && UP.Count > MaxTripCount) 1044 UP.Count = MaxTripCount; 1045 1046 LLVM_DEBUG(dbgs() << " runtime unrolling with count: " << UP.Count 1047 << "\n"); 1048 if (UP.Count < 2) 1049 UP.Count = 0; 1050 return ExplicitUnroll; 1051 } 1052 1053 static LoopUnrollResult tryToUnrollLoop( 1054 Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE, 1055 const TargetTransformInfo &TTI, AssumptionCache &AC, 1056 OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI, 1057 ProfileSummaryInfo *PSI, bool PreserveLCSSA, int OptLevel, 1058 bool OnlyWhenForced, bool ForgetAllSCEV, Optional<unsigned> ProvidedCount, 1059 Optional<unsigned> ProvidedThreshold, Optional<bool> ProvidedAllowPartial, 1060 Optional<bool> ProvidedRuntime, Optional<bool> ProvidedUpperBound, 1061 Optional<bool> ProvidedAllowPeeling, 1062 Optional<bool> ProvidedAllowProfileBasedPeeling, 1063 Optional<unsigned> ProvidedFullUnrollMaxCount) { 1064 LLVM_DEBUG(dbgs() << "Loop Unroll: F[" 1065 << L->getHeader()->getParent()->getName() << "] Loop %" 1066 << L->getHeader()->getName() << "\n"); 1067 TransformationMode TM = hasUnrollTransformation(L); 1068 if (TM & TM_Disable) 1069 return LoopUnrollResult::Unmodified; 1070 if (!L->isLoopSimplifyForm()) { 1071 LLVM_DEBUG( 1072 dbgs() << " Not unrolling loop which is not in loop-simplify form.\n"); 1073 return LoopUnrollResult::Unmodified; 1074 } 1075 1076 // When automtatic unrolling is disabled, do not unroll unless overridden for 1077 // this loop. 1078 if (OnlyWhenForced && !(TM & TM_Enable)) 1079 return LoopUnrollResult::Unmodified; 1080 1081 bool OptForSize = L->getHeader()->getParent()->hasOptSize(); 1082 unsigned NumInlineCandidates; 1083 bool NotDuplicatable; 1084 bool Convergent; 1085 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences( 1086 L, SE, TTI, BFI, PSI, OptLevel, ProvidedThreshold, ProvidedCount, 1087 ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound, 1088 ProvidedFullUnrollMaxCount); 1089 TargetTransformInfo::PeelingPreferences PP = gatherPeelingPreferences( 1090 L, SE, TTI, ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling); 1091 1092 // Exit early if unrolling is disabled. For OptForSize, we pick the loop size 1093 // as threshold later on. 1094 if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0) && 1095 !OptForSize) 1096 return LoopUnrollResult::Unmodified; 1097 1098 SmallPtrSet<const Value *, 32> EphValues; 1099 CodeMetrics::collectEphemeralValues(L, &AC, EphValues); 1100 1101 unsigned LoopSize = 1102 ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent, 1103 TTI, EphValues, UP.BEInsns); 1104 LLVM_DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 1105 if (NotDuplicatable) { 1106 LLVM_DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable" 1107 << " instructions.\n"); 1108 return LoopUnrollResult::Unmodified; 1109 } 1110 1111 // When optimizing for size, use LoopSize + 1 as threshold (we use < Threshold 1112 // later), to (fully) unroll loops, if it does not increase code size. 1113 if (OptForSize) 1114 UP.Threshold = std::max(UP.Threshold, LoopSize + 1); 1115 1116 if (NumInlineCandidates != 0) { 1117 LLVM_DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 1118 return LoopUnrollResult::Unmodified; 1119 } 1120 1121 // Find trip count and trip multiple if count is not available 1122 unsigned TripCount = 0; 1123 unsigned TripMultiple = 1; 1124 // If there are multiple exiting blocks but one of them is the latch, use the 1125 // latch for the trip count estimation. Otherwise insist on a single exiting 1126 // block for the trip count estimation. 1127 BasicBlock *ExitingBlock = L->getLoopLatch(); 1128 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock)) 1129 ExitingBlock = L->getExitingBlock(); 1130 if (ExitingBlock) { 1131 TripCount = SE.getSmallConstantTripCount(L, ExitingBlock); 1132 TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock); 1133 } 1134 1135 // If the loop contains a convergent operation, the prelude we'd add 1136 // to do the first few instructions before we hit the unrolled loop 1137 // is unsafe -- it adds a control-flow dependency to the convergent 1138 // operation. Therefore restrict remainder loop (try unrollig without). 1139 // 1140 // TODO: This is quite conservative. In practice, convergent_op() 1141 // is likely to be called unconditionally in the loop. In this 1142 // case, the program would be ill-formed (on most architectures) 1143 // unless n were the same on all threads in a thread group. 1144 // Assuming n is the same on all threads, any kind of unrolling is 1145 // safe. But currently llvm's notion of convergence isn't powerful 1146 // enough to express this. 1147 if (Convergent) 1148 UP.AllowRemainder = false; 1149 1150 // Try to find the trip count upper bound if we cannot find the exact trip 1151 // count. 1152 unsigned MaxTripCount = 0; 1153 bool MaxOrZero = false; 1154 if (!TripCount) { 1155 MaxTripCount = SE.getSmallConstantMaxTripCount(L); 1156 MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L); 1157 } 1158 1159 // computeUnrollCount() decides whether it is beneficial to use upper bound to 1160 // fully unroll the loop. 1161 bool UseUpperBound = false; 1162 bool IsCountSetExplicitly = computeUnrollCount( 1163 L, TTI, DT, LI, SE, EphValues, &ORE, TripCount, MaxTripCount, MaxOrZero, 1164 TripMultiple, LoopSize, UP, PP, UseUpperBound); 1165 if (!UP.Count) 1166 return LoopUnrollResult::Unmodified; 1167 // Unroll factor (Count) must be less or equal to TripCount. 1168 if (TripCount && UP.Count > TripCount) 1169 UP.Count = TripCount; 1170 1171 // Save loop properties before it is transformed. 1172 MDNode *OrigLoopID = L->getLoopID(); 1173 1174 // Unroll the loop. 1175 Loop *RemainderLoop = nullptr; 1176 LoopUnrollResult UnrollResult = UnrollLoop( 1177 L, 1178 {UP.Count, TripCount, UP.Force, UP.Runtime, UP.AllowExpensiveTripCount, 1179 UseUpperBound, MaxOrZero, TripMultiple, PP.PeelCount, UP.UnrollRemainder, 1180 ForgetAllSCEV}, 1181 LI, &SE, &DT, &AC, &TTI, &ORE, PreserveLCSSA, &RemainderLoop); 1182 if (UnrollResult == LoopUnrollResult::Unmodified) 1183 return LoopUnrollResult::Unmodified; 1184 1185 if (RemainderLoop) { 1186 Optional<MDNode *> RemainderLoopID = 1187 makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll, 1188 LLVMLoopUnrollFollowupRemainder}); 1189 if (RemainderLoopID.hasValue()) 1190 RemainderLoop->setLoopID(RemainderLoopID.getValue()); 1191 } 1192 1193 if (UnrollResult != LoopUnrollResult::FullyUnrolled) { 1194 Optional<MDNode *> NewLoopID = 1195 makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll, 1196 LLVMLoopUnrollFollowupUnrolled}); 1197 if (NewLoopID.hasValue()) { 1198 L->setLoopID(NewLoopID.getValue()); 1199 1200 // Do not setLoopAlreadyUnrolled if loop attributes have been specified 1201 // explicitly. 1202 return UnrollResult; 1203 } 1204 } 1205 1206 // If loop has an unroll count pragma or unrolled by explicitly set count 1207 // mark loop as unrolled to prevent unrolling beyond that requested. 1208 // If the loop was peeled, we already "used up" the profile information 1209 // we had, so we don't want to unroll or peel again. 1210 if (UnrollResult != LoopUnrollResult::FullyUnrolled && 1211 (IsCountSetExplicitly || (PP.PeelProfiledIterations && PP.PeelCount))) 1212 L->setLoopAlreadyUnrolled(); 1213 1214 return UnrollResult; 1215 } 1216 1217 namespace { 1218 1219 class LoopUnroll : public LoopPass { 1220 public: 1221 static char ID; // Pass ID, replacement for typeid 1222 1223 int OptLevel; 1224 1225 /// If false, use a cost model to determine whether unrolling of a loop is 1226 /// profitable. If true, only loops that explicitly request unrolling via 1227 /// metadata are considered. All other loops are skipped. 1228 bool OnlyWhenForced; 1229 1230 /// If false, when SCEV is invalidated, only forget everything in the 1231 /// top-most loop (call forgetTopMostLoop), of the loop being processed. 1232 /// Otherwise, forgetAllLoops and rebuild when needed next. 1233 bool ForgetAllSCEV; 1234 1235 Optional<unsigned> ProvidedCount; 1236 Optional<unsigned> ProvidedThreshold; 1237 Optional<bool> ProvidedAllowPartial; 1238 Optional<bool> ProvidedRuntime; 1239 Optional<bool> ProvidedUpperBound; 1240 Optional<bool> ProvidedAllowPeeling; 1241 Optional<bool> ProvidedAllowProfileBasedPeeling; 1242 Optional<unsigned> ProvidedFullUnrollMaxCount; 1243 1244 LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false, 1245 bool ForgetAllSCEV = false, Optional<unsigned> Threshold = None, 1246 Optional<unsigned> Count = None, 1247 Optional<bool> AllowPartial = None, Optional<bool> Runtime = None, 1248 Optional<bool> UpperBound = None, 1249 Optional<bool> AllowPeeling = None, 1250 Optional<bool> AllowProfileBasedPeeling = None, 1251 Optional<unsigned> ProvidedFullUnrollMaxCount = None) 1252 : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced), 1253 ForgetAllSCEV(ForgetAllSCEV), ProvidedCount(std::move(Count)), 1254 ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial), 1255 ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound), 1256 ProvidedAllowPeeling(AllowPeeling), 1257 ProvidedAllowProfileBasedPeeling(AllowProfileBasedPeeling), 1258 ProvidedFullUnrollMaxCount(ProvidedFullUnrollMaxCount) { 1259 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 1260 } 1261 1262 bool runOnLoop(Loop *L, LPPassManager &LPM) override { 1263 if (skipLoop(L)) 1264 return false; 1265 1266 Function &F = *L->getHeader()->getParent(); 1267 1268 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1269 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1270 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 1271 const TargetTransformInfo &TTI = 1272 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 1273 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1274 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis 1275 // pass. Function analyses need to be preserved across loop transformations 1276 // but ORE cannot be preserved (see comment before the pass definition). 1277 OptimizationRemarkEmitter ORE(&F); 1278 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID); 1279 1280 LoopUnrollResult Result = tryToUnrollLoop( 1281 L, DT, LI, SE, TTI, AC, ORE, nullptr, nullptr, PreserveLCSSA, OptLevel, 1282 OnlyWhenForced, ForgetAllSCEV, ProvidedCount, ProvidedThreshold, 1283 ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound, 1284 ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling, 1285 ProvidedFullUnrollMaxCount); 1286 1287 if (Result == LoopUnrollResult::FullyUnrolled) 1288 LPM.markLoopAsDeleted(*L); 1289 1290 return Result != LoopUnrollResult::Unmodified; 1291 } 1292 1293 /// This transformation requires natural loop information & requires that 1294 /// loop preheaders be inserted into the CFG... 1295 void getAnalysisUsage(AnalysisUsage &AU) const override { 1296 AU.addRequired<AssumptionCacheTracker>(); 1297 AU.addRequired<TargetTransformInfoWrapperPass>(); 1298 // FIXME: Loop passes are required to preserve domtree, and for now we just 1299 // recreate dom info if anything gets unrolled. 1300 getLoopAnalysisUsage(AU); 1301 } 1302 }; 1303 1304 } // end anonymous namespace 1305 1306 char LoopUnroll::ID = 0; 1307 1308 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 1309 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1310 INITIALIZE_PASS_DEPENDENCY(LoopPass) 1311 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 1312 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 1313 1314 Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced, 1315 bool ForgetAllSCEV, int Threshold, int Count, 1316 int AllowPartial, int Runtime, int UpperBound, 1317 int AllowPeeling) { 1318 // TODO: It would make more sense for this function to take the optionals 1319 // directly, but that's dangerous since it would silently break out of tree 1320 // callers. 1321 return new LoopUnroll( 1322 OptLevel, OnlyWhenForced, ForgetAllSCEV, 1323 Threshold == -1 ? None : Optional<unsigned>(Threshold), 1324 Count == -1 ? None : Optional<unsigned>(Count), 1325 AllowPartial == -1 ? None : Optional<bool>(AllowPartial), 1326 Runtime == -1 ? None : Optional<bool>(Runtime), 1327 UpperBound == -1 ? None : Optional<bool>(UpperBound), 1328 AllowPeeling == -1 ? None : Optional<bool>(AllowPeeling)); 1329 } 1330 1331 Pass *llvm::createSimpleLoopUnrollPass(int OptLevel, bool OnlyWhenForced, 1332 bool ForgetAllSCEV) { 1333 return createLoopUnrollPass(OptLevel, OnlyWhenForced, ForgetAllSCEV, -1, -1, 1334 0, 0, 0, 0); 1335 } 1336 1337 PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM, 1338 LoopStandardAnalysisResults &AR, 1339 LPMUpdater &Updater) { 1340 // For the new PM, we can't use OptimizationRemarkEmitter as an analysis 1341 // pass. Function analyses need to be preserved across loop transformations 1342 // but ORE cannot be preserved (see comment before the pass definition). 1343 OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); 1344 1345 // Keep track of the previous loop structure so we can identify new loops 1346 // created by unrolling. 1347 Loop *ParentL = L.getParentLoop(); 1348 SmallPtrSet<Loop *, 4> OldLoops; 1349 if (ParentL) 1350 OldLoops.insert(ParentL->begin(), ParentL->end()); 1351 else 1352 OldLoops.insert(AR.LI.begin(), AR.LI.end()); 1353 1354 std::string LoopName = std::string(L.getName()); 1355 1356 bool Changed = tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, ORE, 1357 /*BFI*/ nullptr, /*PSI*/ nullptr, 1358 /*PreserveLCSSA*/ true, OptLevel, 1359 OnlyWhenForced, ForgetSCEV, /*Count*/ None, 1360 /*Threshold*/ None, /*AllowPartial*/ false, 1361 /*Runtime*/ false, /*UpperBound*/ false, 1362 /*AllowPeeling*/ false, 1363 /*AllowProfileBasedPeeling*/ false, 1364 /*FullUnrollMaxCount*/ None) != 1365 LoopUnrollResult::Unmodified; 1366 if (!Changed) 1367 return PreservedAnalyses::all(); 1368 1369 // The parent must not be damaged by unrolling! 1370 #ifndef NDEBUG 1371 if (ParentL) 1372 ParentL->verifyLoop(); 1373 #endif 1374 1375 // Unrolling can do several things to introduce new loops into a loop nest: 1376 // - Full unrolling clones child loops within the current loop but then 1377 // removes the current loop making all of the children appear to be new 1378 // sibling loops. 1379 // 1380 // When a new loop appears as a sibling loop after fully unrolling, 1381 // its nesting structure has fundamentally changed and we want to revisit 1382 // it to reflect that. 1383 // 1384 // When unrolling has removed the current loop, we need to tell the 1385 // infrastructure that it is gone. 1386 // 1387 // Finally, we support a debugging/testing mode where we revisit child loops 1388 // as well. These are not expected to require further optimizations as either 1389 // they or the loop they were cloned from have been directly visited already. 1390 // But the debugging mode allows us to check this assumption. 1391 bool IsCurrentLoopValid = false; 1392 SmallVector<Loop *, 4> SibLoops; 1393 if (ParentL) 1394 SibLoops.append(ParentL->begin(), ParentL->end()); 1395 else 1396 SibLoops.append(AR.LI.begin(), AR.LI.end()); 1397 erase_if(SibLoops, [&](Loop *SibLoop) { 1398 if (SibLoop == &L) { 1399 IsCurrentLoopValid = true; 1400 return true; 1401 } 1402 1403 // Otherwise erase the loop from the list if it was in the old loops. 1404 return OldLoops.count(SibLoop) != 0; 1405 }); 1406 Updater.addSiblingLoops(SibLoops); 1407 1408 if (!IsCurrentLoopValid) { 1409 Updater.markLoopAsDeleted(L, LoopName); 1410 } else { 1411 // We can only walk child loops if the current loop remained valid. 1412 if (UnrollRevisitChildLoops) { 1413 // Walk *all* of the child loops. 1414 SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end()); 1415 Updater.addChildLoops(ChildLoops); 1416 } 1417 } 1418 1419 return getLoopPassPreservedAnalyses(); 1420 } 1421 1422 PreservedAnalyses LoopUnrollPass::run(Function &F, 1423 FunctionAnalysisManager &AM) { 1424 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 1425 auto &LI = AM.getResult<LoopAnalysis>(F); 1426 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 1427 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1428 auto &AC = AM.getResult<AssumptionAnalysis>(F); 1429 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); 1430 1431 LoopAnalysisManager *LAM = nullptr; 1432 if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(F)) 1433 LAM = &LAMProxy->getManager(); 1434 1435 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); 1436 ProfileSummaryInfo *PSI = 1437 MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); 1438 auto *BFI = (PSI && PSI->hasProfileSummary()) ? 1439 &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr; 1440 1441 bool Changed = false; 1442 1443 // The unroller requires loops to be in simplified form, and also needs LCSSA. 1444 // Since simplification may add new inner loops, it has to run before the 1445 // legality and profitability checks. This means running the loop unroller 1446 // will simplify all loops, regardless of whether anything end up being 1447 // unrolled. 1448 for (auto &L : LI) { 1449 Changed |= 1450 simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */); 1451 Changed |= formLCSSARecursively(*L, DT, &LI, &SE); 1452 } 1453 1454 // Add the loop nests in the reverse order of LoopInfo. See method 1455 // declaration. 1456 SmallPriorityWorklist<Loop *, 4> Worklist; 1457 appendLoopsToWorklist(LI, Worklist); 1458 1459 while (!Worklist.empty()) { 1460 // Because the LoopInfo stores the loops in RPO, we walk the worklist 1461 // from back to front so that we work forward across the CFG, which 1462 // for unrolling is only needed to get optimization remarks emitted in 1463 // a forward order. 1464 Loop &L = *Worklist.pop_back_val(); 1465 #ifndef NDEBUG 1466 Loop *ParentL = L.getParentLoop(); 1467 #endif 1468 1469 // Check if the profile summary indicates that the profiled application 1470 // has a huge working set size, in which case we disable peeling to avoid 1471 // bloating it further. 1472 Optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling; 1473 if (PSI && PSI->hasHugeWorkingSetSize()) 1474 LocalAllowPeeling = false; 1475 std::string LoopName = std::string(L.getName()); 1476 // The API here is quite complex to call and we allow to select some 1477 // flavors of unrolling during construction time (by setting UnrollOpts). 1478 LoopUnrollResult Result = tryToUnrollLoop( 1479 &L, DT, &LI, SE, TTI, AC, ORE, BFI, PSI, 1480 /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced, 1481 UnrollOpts.ForgetSCEV, /*Count*/ None, 1482 /*Threshold*/ None, UnrollOpts.AllowPartial, UnrollOpts.AllowRuntime, 1483 UnrollOpts.AllowUpperBound, LocalAllowPeeling, 1484 UnrollOpts.AllowProfileBasedPeeling, UnrollOpts.FullUnrollMaxCount); 1485 Changed |= Result != LoopUnrollResult::Unmodified; 1486 1487 // The parent must not be damaged by unrolling! 1488 #ifndef NDEBUG 1489 if (Result != LoopUnrollResult::Unmodified && ParentL) 1490 ParentL->verifyLoop(); 1491 #endif 1492 1493 // Clear any cached analysis results for L if we removed it completely. 1494 if (LAM && Result == LoopUnrollResult::FullyUnrolled) 1495 LAM->clear(L, LoopName); 1496 } 1497 1498 if (!Changed) 1499 return PreservedAnalyses::all(); 1500 1501 return getLoopPassPreservedAnalyses(); 1502 } 1503