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