1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass implements a simple loop unroller. It works best when loops have 11 // been canonicalized by the -indvars pass, allowing it to determine the trip 12 // counts of loops easily. 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Scalar.h" 16 #include "llvm/ADT/SetVector.h" 17 #include "llvm/Analysis/GlobalsModRef.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/CodeMetrics.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/LoopPass.h" 22 #include "llvm/Analysis/LoopUnrollAnalyzer.h" 23 #include "llvm/Analysis/ScalarEvolution.h" 24 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 25 #include "llvm/Analysis/TargetTransformInfo.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/DiagnosticInfo.h" 28 #include "llvm/IR/Dominators.h" 29 #include "llvm/IR/InstVisitor.h" 30 #include "llvm/IR/IntrinsicInst.h" 31 #include "llvm/IR/Metadata.h" 32 #include "llvm/Support/CommandLine.h" 33 #include "llvm/Support/Debug.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include "llvm/Transforms/Utils/UnrollLoop.h" 36 #include <climits> 37 38 using namespace llvm; 39 40 #define DEBUG_TYPE "loop-unroll" 41 42 static cl::opt<unsigned> 43 UnrollThreshold("unroll-threshold", cl::Hidden, 44 cl::desc("The baseline cost threshold for loop unrolling")); 45 46 static cl::opt<unsigned> UnrollPercentDynamicCostSavedThreshold( 47 "unroll-percent-dynamic-cost-saved-threshold", cl::Hidden, 48 cl::desc("The percentage of estimated dynamic cost which must be saved by " 49 "unrolling to allow unrolling up to the max threshold.")); 50 51 static cl::opt<unsigned> UnrollDynamicCostSavingsDiscount( 52 "unroll-dynamic-cost-savings-discount", cl::Hidden, 53 cl::desc("This is the amount discounted from the total unroll cost when " 54 "the unrolled form has a high dynamic cost savings (triggered by " 55 "the '-unroll-perecent-dynamic-cost-saved-threshold' flag).")); 56 57 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze( 58 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden, 59 cl::desc("Don't allow loop unrolling to simulate more than this number of" 60 "iterations when checking full unroll profitability")); 61 62 static cl::opt<unsigned> 63 UnrollCount("unroll-count", cl::Hidden, 64 cl::desc("Use this unroll count for all loops including those with " 65 "unroll_count pragma values, for testing purposes")); 66 67 static cl::opt<bool> 68 UnrollAllowPartial("unroll-allow-partial", cl::Hidden, 69 cl::desc("Allows loops to be partially unrolled until " 70 "-unroll-threshold loop size is reached.")); 71 72 static cl::opt<bool> 73 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::Hidden, 74 cl::desc("Unroll loops with run-time trip counts")); 75 76 static cl::opt<unsigned> 77 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden, 78 cl::desc("Unrolled size limit for loops with an unroll(full) or " 79 "unroll_count pragma.")); 80 81 82 /// A magic value for use with the Threshold parameter to indicate 83 /// that the loop unroll should be performed regardless of how much 84 /// code expansion would result. 85 static const unsigned NoThreshold = UINT_MAX; 86 87 /// Default unroll count for loops with run-time trip count if 88 /// -unroll-count is not set 89 static const unsigned DefaultUnrollRuntimeCount = 8; 90 91 /// Gather the various unrolling parameters based on the defaults, compiler 92 /// flags, TTI overrides, pragmas, and user specified parameters. 93 static TargetTransformInfo::UnrollingPreferences gatherUnrollingPreferences( 94 Loop *L, const TargetTransformInfo &TTI, Optional<unsigned> UserThreshold, 95 Optional<unsigned> UserCount, Optional<bool> UserAllowPartial, 96 Optional<bool> UserRuntime, unsigned PragmaCount, bool PragmaFullUnroll, 97 bool PragmaEnableUnroll, unsigned TripCount) { 98 TargetTransformInfo::UnrollingPreferences UP; 99 100 // Set up the defaults 101 UP.Threshold = 150; 102 UP.PercentDynamicCostSavedThreshold = 20; 103 UP.DynamicCostSavingsDiscount = 2000; 104 UP.OptSizeThreshold = 50; 105 UP.PartialThreshold = UP.Threshold; 106 UP.PartialOptSizeThreshold = UP.OptSizeThreshold; 107 UP.Count = 0; 108 UP.MaxCount = UINT_MAX; 109 UP.Partial = false; 110 UP.Runtime = false; 111 UP.AllowExpensiveTripCount = false; 112 113 // Override with any target specific settings 114 TTI.getUnrollingPreferences(L, UP); 115 116 // Apply size attributes 117 if (L->getHeader()->getParent()->optForSize()) { 118 UP.Threshold = UP.OptSizeThreshold; 119 UP.PartialThreshold = UP.PartialOptSizeThreshold; 120 } 121 122 // Apply unroll count pragmas 123 if (PragmaCount) 124 UP.Count = PragmaCount; 125 else if (PragmaFullUnroll) 126 UP.Count = TripCount; 127 128 // Apply any user values specified by cl::opt 129 if (UnrollThreshold.getNumOccurrences() > 0) { 130 UP.Threshold = UnrollThreshold; 131 UP.PartialThreshold = UnrollThreshold; 132 } 133 if (UnrollPercentDynamicCostSavedThreshold.getNumOccurrences() > 0) 134 UP.PercentDynamicCostSavedThreshold = 135 UnrollPercentDynamicCostSavedThreshold; 136 if (UnrollDynamicCostSavingsDiscount.getNumOccurrences() > 0) 137 UP.DynamicCostSavingsDiscount = UnrollDynamicCostSavingsDiscount; 138 if (UnrollCount.getNumOccurrences() > 0) 139 UP.Count = UnrollCount; 140 if (UnrollAllowPartial.getNumOccurrences() > 0) 141 UP.Partial = UnrollAllowPartial; 142 if (UnrollRuntime.getNumOccurrences() > 0) 143 UP.Runtime = UnrollRuntime; 144 145 // Apply user values provided by argument 146 if (UserThreshold.hasValue()) { 147 UP.Threshold = *UserThreshold; 148 UP.PartialThreshold = *UserThreshold; 149 } 150 if (UserCount.hasValue()) 151 UP.Count = *UserCount; 152 if (UserAllowPartial.hasValue()) 153 UP.Partial = *UserAllowPartial; 154 if (UserRuntime.hasValue()) 155 UP.Runtime = *UserRuntime; 156 157 if (PragmaCount > 0 || 158 ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount != 0)) { 159 // If the loop has an unrolling pragma, we want to be more aggressive with 160 // unrolling limits. Set thresholds to at least the PragmaTheshold value 161 // which is larger than the default limits. 162 if (UP.Threshold != NoThreshold) 163 UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold); 164 if (UP.PartialThreshold != NoThreshold) 165 UP.PartialThreshold = 166 std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold); 167 } 168 169 return UP; 170 } 171 172 namespace { 173 struct EstimatedUnrollCost { 174 /// \brief The estimated cost after unrolling. 175 int UnrolledCost; 176 177 /// \brief The estimated dynamic cost of executing the instructions in the 178 /// rolled form. 179 int RolledDynamicCost; 180 }; 181 } 182 183 /// \brief Figure out if the loop is worth full unrolling. 184 /// 185 /// Complete loop unrolling can make some loads constant, and we need to know 186 /// if that would expose any further optimization opportunities. This routine 187 /// estimates this optimization. It computes cost of unrolled loop 188 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By 189 /// dynamic cost we mean that we won't count costs of blocks that are known not 190 /// to be executed (i.e. if we have a branch in the loop and we know that at the 191 /// given iteration its condition would be resolved to true, we won't add up the 192 /// cost of the 'false'-block). 193 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If 194 /// the analysis failed (no benefits expected from the unrolling, or the loop is 195 /// too big to analyze), the returned value is None. 196 static Optional<EstimatedUnrollCost> 197 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, DominatorTree &DT, 198 ScalarEvolution &SE, const TargetTransformInfo &TTI, 199 int MaxUnrolledLoopSize) { 200 // We want to be able to scale offsets by the trip count and add more offsets 201 // to them without checking for overflows, and we already don't want to 202 // analyze *massive* trip counts, so we force the max to be reasonably small. 203 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) && 204 "The unroll iterations max is too large!"); 205 206 // Don't simulate loops with a big or unknown tripcount 207 if (!UnrollMaxIterationsCountToAnalyze || !TripCount || 208 TripCount > UnrollMaxIterationsCountToAnalyze) 209 return None; 210 211 SmallSetVector<BasicBlock *, 16> BBWorklist; 212 DenseMap<Value *, Constant *> SimplifiedValues; 213 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues; 214 215 // The estimated cost of the unrolled form of the loop. We try to estimate 216 // this by simplifying as much as we can while computing the estimate. 217 int UnrolledCost = 0; 218 // We also track the estimated dynamic (that is, actually executed) cost in 219 // the rolled form. This helps identify cases when the savings from unrolling 220 // aren't just exposing dead control flows, but actual reduced dynamic 221 // instructions due to the simplifications which we expect to occur after 222 // unrolling. 223 int RolledDynamicCost = 0; 224 225 // Ensure that we don't violate the loop structure invariants relied on by 226 // this analysis. 227 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first."); 228 assert(L->isLCSSAForm(DT) && 229 "Must have loops in LCSSA form to track live-out values."); 230 231 DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n"); 232 233 // Simulate execution of each iteration of the loop counting instructions, 234 // which would be simplified. 235 // Since the same load will take different values on different iterations, 236 // we literally have to go through all loop's iterations. 237 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) { 238 DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n"); 239 240 // Prepare for the iteration by collecting any simplified entry or backedge 241 // inputs. 242 for (Instruction &I : *L->getHeader()) { 243 auto *PHI = dyn_cast<PHINode>(&I); 244 if (!PHI) 245 break; 246 247 // The loop header PHI nodes must have exactly two input: one from the 248 // loop preheader and one from the loop latch. 249 assert( 250 PHI->getNumIncomingValues() == 2 && 251 "Must have an incoming value only for the preheader and the latch."); 252 253 Value *V = PHI->getIncomingValueForBlock( 254 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch()); 255 Constant *C = dyn_cast<Constant>(V); 256 if (Iteration != 0 && !C) 257 C = SimplifiedValues.lookup(V); 258 if (C) 259 SimplifiedInputValues.push_back({PHI, C}); 260 } 261 262 // Now clear and re-populate the map for the next iteration. 263 SimplifiedValues.clear(); 264 while (!SimplifiedInputValues.empty()) 265 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val()); 266 267 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE); 268 269 BBWorklist.clear(); 270 BBWorklist.insert(L->getHeader()); 271 // Note that we *must not* cache the size, this loop grows the worklist. 272 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) { 273 BasicBlock *BB = BBWorklist[Idx]; 274 275 // Visit all instructions in the given basic block and try to simplify 276 // it. We don't change the actual IR, just count optimization 277 // opportunities. 278 for (Instruction &I : *BB) { 279 int InstCost = TTI.getUserCost(&I); 280 281 // Visit the instruction to analyze its loop cost after unrolling, 282 // and if the visitor returns false, include this instruction in the 283 // unrolled cost. 284 if (!Analyzer.visit(I)) 285 UnrolledCost += InstCost; 286 else { 287 DEBUG(dbgs() << " " << I 288 << " would be simplified if loop is unrolled.\n"); 289 (void)0; 290 } 291 292 // Also track this instructions expected cost when executing the rolled 293 // loop form. 294 RolledDynamicCost += InstCost; 295 296 // If unrolled body turns out to be too big, bail out. 297 if (UnrolledCost > MaxUnrolledLoopSize) { 298 DEBUG(dbgs() << " Exceeded threshold.. exiting.\n" 299 << " UnrolledCost: " << UnrolledCost 300 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize 301 << "\n"); 302 return None; 303 } 304 } 305 306 TerminatorInst *TI = BB->getTerminator(); 307 308 // Add in the live successors by first checking whether we have terminator 309 // that may be simplified based on the values simplified by this call. 310 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 311 if (BI->isConditional()) { 312 if (Constant *SimpleCond = 313 SimplifiedValues.lookup(BI->getCondition())) { 314 BasicBlock *Succ = nullptr; 315 // Just take the first successor if condition is undef 316 if (isa<UndefValue>(SimpleCond)) 317 Succ = BI->getSuccessor(0); 318 else 319 Succ = BI->getSuccessor( 320 cast<ConstantInt>(SimpleCond)->isZero() ? 1 : 0); 321 if (L->contains(Succ)) 322 BBWorklist.insert(Succ); 323 continue; 324 } 325 } 326 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 327 if (Constant *SimpleCond = 328 SimplifiedValues.lookup(SI->getCondition())) { 329 BasicBlock *Succ = nullptr; 330 // Just take the first successor if condition is undef 331 if (isa<UndefValue>(SimpleCond)) 332 Succ = SI->getSuccessor(0); 333 else 334 Succ = SI->findCaseValue(cast<ConstantInt>(SimpleCond)) 335 .getCaseSuccessor(); 336 if (L->contains(Succ)) 337 BBWorklist.insert(Succ); 338 continue; 339 } 340 } 341 342 // Add BB's successors to the worklist. 343 for (BasicBlock *Succ : successors(BB)) 344 if (L->contains(Succ)) 345 BBWorklist.insert(Succ); 346 } 347 348 // If we found no optimization opportunities on the first iteration, we 349 // won't find them on later ones too. 350 if (UnrolledCost == RolledDynamicCost) { 351 DEBUG(dbgs() << " No opportunities found.. exiting.\n" 352 << " UnrolledCost: " << UnrolledCost << "\n"); 353 return None; 354 } 355 } 356 DEBUG(dbgs() << "Analysis finished:\n" 357 << "UnrolledCost: " << UnrolledCost << ", " 358 << "RolledDynamicCost: " << RolledDynamicCost << "\n"); 359 return {{UnrolledCost, RolledDynamicCost}}; 360 } 361 362 /// ApproximateLoopSize - Approximate the size of the loop. 363 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, 364 bool &NotDuplicatable, 365 const TargetTransformInfo &TTI, 366 AssumptionCache *AC) { 367 SmallPtrSet<const Value *, 32> EphValues; 368 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 369 370 CodeMetrics Metrics; 371 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 372 I != E; ++I) 373 Metrics.analyzeBasicBlock(*I, TTI, EphValues); 374 NumCalls = Metrics.NumInlineCandidates; 375 NotDuplicatable = Metrics.notDuplicatable; 376 377 unsigned LoopSize = Metrics.NumInsts; 378 379 // Don't allow an estimate of size zero. This would allows unrolling of loops 380 // with huge iteration counts, which is a compile time problem even if it's 381 // not a problem for code quality. Also, the code using this size may assume 382 // that each loop has at least three instructions (likely a conditional 383 // branch, a comparison feeding that branch, and some kind of loop increment 384 // feeding that comparison instruction). 385 LoopSize = std::max(LoopSize, 3u); 386 387 return LoopSize; 388 } 389 390 // Returns the loop hint metadata node with the given name (for example, 391 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is 392 // returned. 393 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) { 394 if (MDNode *LoopID = L->getLoopID()) 395 return GetUnrollMetadata(LoopID, Name); 396 return nullptr; 397 } 398 399 // Returns true if the loop has an unroll(full) pragma. 400 static bool HasUnrollFullPragma(const Loop *L) { 401 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full"); 402 } 403 404 // Returns true if the loop has an unroll(enable) pragma. This metadata is used 405 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives. 406 static bool HasUnrollEnablePragma(const Loop *L) { 407 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable"); 408 } 409 410 // Returns true if the loop has an unroll(disable) pragma. 411 static bool HasUnrollDisablePragma(const Loop *L) { 412 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable"); 413 } 414 415 // Returns true if the loop has an runtime unroll(disable) pragma. 416 static bool HasRuntimeUnrollDisablePragma(const Loop *L) { 417 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable"); 418 } 419 420 // If loop has an unroll_count pragma return the (necessarily 421 // positive) value from the pragma. Otherwise return 0. 422 static unsigned UnrollCountPragmaValue(const Loop *L) { 423 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count"); 424 if (MD) { 425 assert(MD->getNumOperands() == 2 && 426 "Unroll count hint metadata should have two operands."); 427 unsigned Count = 428 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 429 assert(Count >= 1 && "Unroll count must be positive."); 430 return Count; 431 } 432 return 0; 433 } 434 435 // Remove existing unroll metadata and add unroll disable metadata to 436 // indicate the loop has already been unrolled. This prevents a loop 437 // from being unrolled more than is directed by a pragma if the loop 438 // unrolling pass is run more than once (which it generally is). 439 static void SetLoopAlreadyUnrolled(Loop *L) { 440 MDNode *LoopID = L->getLoopID(); 441 if (!LoopID) return; 442 443 // First remove any existing loop unrolling metadata. 444 SmallVector<Metadata *, 4> MDs; 445 // Reserve first location for self reference to the LoopID metadata node. 446 MDs.push_back(nullptr); 447 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 448 bool IsUnrollMetadata = false; 449 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 450 if (MD) { 451 const MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 452 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll."); 453 } 454 if (!IsUnrollMetadata) 455 MDs.push_back(LoopID->getOperand(i)); 456 } 457 458 // Add unroll(disable) metadata to disable future unrolling. 459 LLVMContext &Context = L->getHeader()->getContext(); 460 SmallVector<Metadata *, 1> DisableOperands; 461 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable")); 462 MDNode *DisableNode = MDNode::get(Context, DisableOperands); 463 MDs.push_back(DisableNode); 464 465 MDNode *NewLoopID = MDNode::get(Context, MDs); 466 // Set operand 0 to refer to the loop id itself. 467 NewLoopID->replaceOperandWith(0, NewLoopID); 468 L->setLoopID(NewLoopID); 469 } 470 471 static bool canUnrollCompletely(Loop *L, unsigned Threshold, 472 unsigned PercentDynamicCostSavedThreshold, 473 unsigned DynamicCostSavingsDiscount, 474 uint64_t UnrolledCost, 475 uint64_t RolledDynamicCost) { 476 if (Threshold == NoThreshold) { 477 DEBUG(dbgs() << " Can fully unroll, because no threshold is set.\n"); 478 return true; 479 } 480 481 if (UnrolledCost <= Threshold) { 482 DEBUG(dbgs() << " Can fully unroll, because unrolled cost: " 483 << UnrolledCost << "<" << Threshold << "\n"); 484 return true; 485 } 486 487 assert(UnrolledCost && "UnrolledCost can't be 0 at this point."); 488 assert(RolledDynamicCost >= UnrolledCost && 489 "Cannot have a higher unrolled cost than a rolled cost!"); 490 491 // Compute the percentage of the dynamic cost in the rolled form that is 492 // saved when unrolled. If unrolling dramatically reduces the estimated 493 // dynamic cost of the loop, we use a higher threshold to allow more 494 // unrolling. 495 unsigned PercentDynamicCostSaved = 496 (uint64_t)(RolledDynamicCost - UnrolledCost) * 100ull / RolledDynamicCost; 497 498 if (PercentDynamicCostSaved >= PercentDynamicCostSavedThreshold && 499 (int64_t)UnrolledCost - (int64_t)DynamicCostSavingsDiscount <= 500 (int64_t)Threshold) { 501 DEBUG(dbgs() << " Can fully unroll, because unrolling will reduce the " 502 "expected dynamic cost by " << PercentDynamicCostSaved 503 << "% (threshold: " << PercentDynamicCostSavedThreshold 504 << "%)\n" 505 << " and the unrolled cost (" << UnrolledCost 506 << ") is less than the max threshold (" 507 << DynamicCostSavingsDiscount << ").\n"); 508 return true; 509 } 510 511 DEBUG(dbgs() << " Too large to fully unroll:\n"); 512 DEBUG(dbgs() << " Threshold: " << Threshold << "\n"); 513 DEBUG(dbgs() << " Max threshold: " << DynamicCostSavingsDiscount << "\n"); 514 DEBUG(dbgs() << " Percent cost saved threshold: " 515 << PercentDynamicCostSavedThreshold << "%\n"); 516 DEBUG(dbgs() << " Unrolled cost: " << UnrolledCost << "\n"); 517 DEBUG(dbgs() << " Rolled dynamic cost: " << RolledDynamicCost << "\n"); 518 DEBUG(dbgs() << " Percent cost saved: " << PercentDynamicCostSaved 519 << "\n"); 520 return false; 521 } 522 523 static bool tryToUnrollLoop(Loop *L, DominatorTree &DT, LoopInfo *LI, 524 ScalarEvolution *SE, const TargetTransformInfo &TTI, 525 AssumptionCache &AC, bool PreserveLCSSA, 526 Optional<unsigned> ProvidedCount, 527 Optional<unsigned> ProvidedThreshold, 528 Optional<bool> ProvidedAllowPartial, 529 Optional<bool> ProvidedRuntime) { 530 BasicBlock *Header = L->getHeader(); 531 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 532 << "] Loop %" << Header->getName() << "\n"); 533 534 if (HasUnrollDisablePragma(L)) { 535 return false; 536 } 537 bool PragmaFullUnroll = HasUnrollFullPragma(L); 538 bool PragmaEnableUnroll = HasUnrollEnablePragma(L); 539 unsigned PragmaCount = UnrollCountPragmaValue(L); 540 bool HasPragma = PragmaFullUnroll || PragmaEnableUnroll || PragmaCount > 0; 541 542 // Find trip count and trip multiple if count is not available 543 unsigned TripCount = 0; 544 unsigned TripMultiple = 1; 545 // If there are multiple exiting blocks but one of them is the latch, use the 546 // latch for the trip count estimation. Otherwise insist on a single exiting 547 // block for the trip count estimation. 548 BasicBlock *ExitingBlock = L->getLoopLatch(); 549 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock)) 550 ExitingBlock = L->getExitingBlock(); 551 if (ExitingBlock) { 552 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock); 553 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock); 554 } 555 556 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences( 557 L, TTI, ProvidedThreshold, ProvidedCount, ProvidedAllowPartial, 558 ProvidedRuntime, PragmaCount, PragmaFullUnroll, PragmaEnableUnroll, 559 TripCount); 560 561 unsigned Count = UP.Count; 562 bool CountSetExplicitly = Count != 0; 563 // Use a heuristic count if we didn't set anything explicitly. 564 if (!CountSetExplicitly) 565 Count = TripCount == 0 ? DefaultUnrollRuntimeCount : TripCount; 566 if (TripCount && Count > TripCount) 567 Count = TripCount; 568 569 unsigned NumInlineCandidates; 570 bool notDuplicatable; 571 unsigned LoopSize = 572 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC); 573 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 574 575 // When computing the unrolled size, note that the conditional branch on the 576 // backedge and the comparison feeding it are not replicated like the rest of 577 // the loop body (which is why 2 is subtracted). 578 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2; 579 if (notDuplicatable) { 580 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable" 581 << " instructions.\n"); 582 return false; 583 } 584 if (NumInlineCandidates != 0) { 585 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 586 return false; 587 } 588 589 // Given Count, TripCount and thresholds determine the type of 590 // unrolling which is to be performed. 591 enum { Full = 0, Partial = 1, Runtime = 2 }; 592 int Unrolling; 593 if (TripCount && Count == TripCount) { 594 Unrolling = Partial; 595 // If the loop is really small, we don't need to run an expensive analysis. 596 if (canUnrollCompletely(L, UP.Threshold, 100, UP.DynamicCostSavingsDiscount, 597 UnrolledSize, UnrolledSize)) { 598 Unrolling = Full; 599 } else { 600 // The loop isn't that small, but we still can fully unroll it if that 601 // helps to remove a significant number of instructions. 602 // To check that, run additional analysis on the loop. 603 if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost( 604 L, TripCount, DT, *SE, TTI, 605 UP.Threshold + UP.DynamicCostSavingsDiscount)) 606 if (canUnrollCompletely(L, UP.Threshold, 607 UP.PercentDynamicCostSavedThreshold, 608 UP.DynamicCostSavingsDiscount, 609 Cost->UnrolledCost, Cost->RolledDynamicCost)) { 610 Unrolling = Full; 611 } 612 } 613 } else if (TripCount && Count < TripCount) { 614 Unrolling = Partial; 615 } else { 616 Unrolling = Runtime; 617 } 618 619 // Reduce count based on the type of unrolling and the threshold values. 620 unsigned OriginalCount = Count; 621 bool AllowRuntime = PragmaEnableUnroll || (PragmaCount > 0) || UP.Runtime; 622 // Don't unroll a runtime trip count loop with unroll full pragma. 623 if (HasRuntimeUnrollDisablePragma(L) || PragmaFullUnroll) { 624 AllowRuntime = false; 625 } 626 if (Unrolling == Partial) { 627 bool AllowPartial = PragmaEnableUnroll || UP.Partial; 628 if (!AllowPartial && !CountSetExplicitly) { 629 DEBUG(dbgs() << " will not try to unroll partially because " 630 << "-unroll-allow-partial not given\n"); 631 return false; 632 } 633 if (UP.PartialThreshold != NoThreshold && 634 UnrolledSize > UP.PartialThreshold) { 635 // Reduce unroll count to be modulo of TripCount for partial unrolling. 636 Count = (std::max(UP.PartialThreshold, 3u) - 2) / (LoopSize - 2); 637 while (Count != 0 && TripCount % Count != 0) 638 Count--; 639 } 640 } else if (Unrolling == Runtime) { 641 if (!AllowRuntime && !CountSetExplicitly) { 642 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count " 643 << "-unroll-runtime not given\n"); 644 return false; 645 } 646 // Reduce unroll count to be the largest power-of-two factor of 647 // the original count which satisfies the threshold limit. 648 while (Count != 0 && UnrolledSize > UP.PartialThreshold) { 649 Count >>= 1; 650 UnrolledSize = (LoopSize-2) * Count + 2; 651 } 652 if (Count > UP.MaxCount) 653 Count = UP.MaxCount; 654 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 655 } 656 657 if (HasPragma) { 658 if (PragmaCount != 0) 659 // If loop has an unroll count pragma mark loop as unrolled to prevent 660 // unrolling beyond that requested by the pragma. 661 SetLoopAlreadyUnrolled(L); 662 663 // Emit optimization remarks if we are unable to unroll the loop 664 // as directed by a pragma. 665 DebugLoc LoopLoc = L->getStartLoc(); 666 Function *F = Header->getParent(); 667 LLVMContext &Ctx = F->getContext(); 668 if ((PragmaCount > 0) && Count != OriginalCount) { 669 emitOptimizationRemarkMissed( 670 Ctx, DEBUG_TYPE, *F, LoopLoc, 671 "Unable to unroll loop the number of times directed by " 672 "unroll_count pragma because unrolled size is too large."); 673 } else if (PragmaFullUnroll && !TripCount) { 674 emitOptimizationRemarkMissed( 675 Ctx, DEBUG_TYPE, *F, LoopLoc, 676 "Unable to fully unroll loop as directed by unroll(full) pragma " 677 "because loop has a runtime trip count."); 678 } else if (PragmaEnableUnroll && Count != TripCount && Count < 2) { 679 emitOptimizationRemarkMissed( 680 Ctx, DEBUG_TYPE, *F, LoopLoc, 681 "Unable to unroll loop as directed by unroll(enable) pragma because " 682 "unrolled size is too large."); 683 } else if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount && 684 Count != TripCount) { 685 emitOptimizationRemarkMissed( 686 Ctx, DEBUG_TYPE, *F, LoopLoc, 687 "Unable to fully unroll loop as directed by unroll pragma because " 688 "unrolled size is too large."); 689 } 690 } 691 692 if (Unrolling != Full && Count < 2) { 693 // Partial unrolling by 1 is a nop. For full unrolling, a factor 694 // of 1 makes sense because loop control can be eliminated. 695 return false; 696 } 697 698 // Unroll the loop. 699 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount, 700 TripMultiple, LI, SE, &DT, &AC, PreserveLCSSA)) 701 return false; 702 703 return true; 704 } 705 706 namespace { 707 class LoopUnroll : public LoopPass { 708 public: 709 static char ID; // Pass ID, replacement for typeid 710 LoopUnroll(Optional<unsigned> Threshold = None, 711 Optional<unsigned> Count = None, 712 Optional<bool> AllowPartial = None, Optional<bool> Runtime = None) 713 : LoopPass(ID), ProvidedCount(Count), ProvidedThreshold(Threshold), 714 ProvidedAllowPartial(AllowPartial), ProvidedRuntime(Runtime) { 715 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 716 } 717 718 Optional<unsigned> ProvidedCount; 719 Optional<unsigned> ProvidedThreshold; 720 Optional<bool> ProvidedAllowPartial; 721 Optional<bool> ProvidedRuntime; 722 723 bool runOnLoop(Loop *L, LPPassManager &) override { 724 if (skipOptnoneFunction(L)) 725 return false; 726 727 Function &F = *L->getHeader()->getParent(); 728 729 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 730 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 731 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 732 const TargetTransformInfo &TTI = 733 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 734 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 735 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID); 736 737 return tryToUnrollLoop(L, DT, LI, SE, TTI, AC, PreserveLCSSA, ProvidedCount, 738 ProvidedThreshold, ProvidedAllowPartial, 739 ProvidedRuntime); 740 } 741 742 /// This transformation requires natural loop information & requires that 743 /// loop preheaders be inserted into the CFG... 744 /// 745 void getAnalysisUsage(AnalysisUsage &AU) const override { 746 AU.addRequired<AssumptionCacheTracker>(); 747 AU.addRequired<DominatorTreeWrapperPass>(); 748 AU.addRequired<LoopInfoWrapperPass>(); 749 AU.addPreserved<LoopInfoWrapperPass>(); 750 AU.addRequiredID(LoopSimplifyID); 751 AU.addPreservedID(LoopSimplifyID); 752 AU.addRequiredID(LCSSAID); 753 AU.addPreservedID(LCSSAID); 754 AU.addRequired<ScalarEvolutionWrapperPass>(); 755 AU.addPreserved<ScalarEvolutionWrapperPass>(); 756 AU.addRequired<TargetTransformInfoWrapperPass>(); 757 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 758 // If loop unroll does not preserve dom info then LCSSA pass on next 759 // loop will receive invalid dom info. 760 // For now, recreate dom info, if loop is unrolled. 761 AU.addPreserved<DominatorTreeWrapperPass>(); 762 AU.addPreserved<GlobalsAAWrapperPass>(); 763 } 764 }; 765 } 766 767 char LoopUnroll::ID = 0; 768 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 769 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 770 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 771 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 772 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 773 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 774 INITIALIZE_PASS_DEPENDENCY(LCSSA) 775 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 776 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 777 778 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial, 779 int Runtime) { 780 // TODO: It would make more sense for this function to take the optionals 781 // directly, but that's dangerous since it would silently break out of tree 782 // callers. 783 return new LoopUnroll(Threshold == -1 ? None : Optional<unsigned>(Threshold), 784 Count == -1 ? None : Optional<unsigned>(Count), 785 AllowPartial == -1 ? None 786 : Optional<bool>(AllowPartial), 787 Runtime == -1 ? None : Optional<bool>(Runtime)); 788 } 789 790 Pass *llvm::createSimpleLoopUnrollPass() { 791 return llvm::createLoopUnrollPass(-1, -1, 0, 0); 792 } 793