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