1 //===- LoopUnrollAndJam.cpp - Loop unroll and jam 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 an unroll and jam pass. Most of the work is done by 10 // Utils/UnrollLoopAndJam.cpp. 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Transforms/Scalar/LoopUnrollAndJamPass.h" 14 #include "llvm/ADT/ArrayRef.h" 15 #include "llvm/ADT/None.h" 16 #include "llvm/ADT/Optional.h" 17 #include "llvm/ADT/PriorityWorklist.h" 18 #include "llvm/ADT/SmallPtrSet.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/Analysis/AssumptionCache.h" 21 #include "llvm/Analysis/CodeMetrics.h" 22 #include "llvm/Analysis/DependenceAnalysis.h" 23 #include "llvm/Analysis/LoopAnalysisManager.h" 24 #include "llvm/Analysis/LoopInfo.h" 25 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 26 #include "llvm/Analysis/ScalarEvolution.h" 27 #include "llvm/Analysis/TargetTransformInfo.h" 28 #include "llvm/IR/BasicBlock.h" 29 #include "llvm/IR/Constants.h" 30 #include "llvm/IR/Dominators.h" 31 #include "llvm/IR/Function.h" 32 #include "llvm/IR/Instructions.h" 33 #include "llvm/IR/Metadata.h" 34 #include "llvm/IR/PassManager.h" 35 #include "llvm/InitializePasses.h" 36 #include "llvm/Pass.h" 37 #include "llvm/PassRegistry.h" 38 #include "llvm/Support/Casting.h" 39 #include "llvm/Support/CommandLine.h" 40 #include "llvm/Support/Compiler.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/raw_ostream.h" 43 #include "llvm/Transforms/Scalar.h" 44 #include "llvm/Transforms/Utils/LoopSimplify.h" 45 #include "llvm/Transforms/Utils/LoopUtils.h" 46 #include "llvm/Transforms/Utils/UnrollLoop.h" 47 #include <cassert> 48 #include <cstdint> 49 #include <vector> 50 51 namespace llvm { 52 class Instruction; 53 class Value; 54 } // namespace llvm 55 56 using namespace llvm; 57 58 #define DEBUG_TYPE "loop-unroll-and-jam" 59 60 /// @{ 61 /// Metadata attribute names 62 static const char *const LLVMLoopUnrollAndJamFollowupAll = 63 "llvm.loop.unroll_and_jam.followup_all"; 64 static const char *const LLVMLoopUnrollAndJamFollowupInner = 65 "llvm.loop.unroll_and_jam.followup_inner"; 66 static const char *const LLVMLoopUnrollAndJamFollowupOuter = 67 "llvm.loop.unroll_and_jam.followup_outer"; 68 static const char *const LLVMLoopUnrollAndJamFollowupRemainderInner = 69 "llvm.loop.unroll_and_jam.followup_remainder_inner"; 70 static const char *const LLVMLoopUnrollAndJamFollowupRemainderOuter = 71 "llvm.loop.unroll_and_jam.followup_remainder_outer"; 72 /// @} 73 74 static cl::opt<bool> 75 AllowUnrollAndJam("allow-unroll-and-jam", cl::Hidden, 76 cl::desc("Allows loops to be unroll-and-jammed.")); 77 78 static cl::opt<unsigned> UnrollAndJamCount( 79 "unroll-and-jam-count", cl::Hidden, 80 cl::desc("Use this unroll count for all loops including those with " 81 "unroll_and_jam_count pragma values, for testing purposes")); 82 83 static cl::opt<unsigned> UnrollAndJamThreshold( 84 "unroll-and-jam-threshold", cl::init(60), cl::Hidden, 85 cl::desc("Threshold to use for inner loop when doing unroll and jam.")); 86 87 static cl::opt<unsigned> PragmaUnrollAndJamThreshold( 88 "pragma-unroll-and-jam-threshold", cl::init(1024), cl::Hidden, 89 cl::desc("Unrolled size limit for loops with an unroll_and_jam(full) or " 90 "unroll_count pragma.")); 91 92 // Returns the loop hint metadata node with the given name (for example, 93 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is 94 // returned. 95 static MDNode *getUnrollMetadataForLoop(const Loop *L, StringRef Name) { 96 if (MDNode *LoopID = L->getLoopID()) 97 return GetUnrollMetadata(LoopID, Name); 98 return nullptr; 99 } 100 101 // Returns true if the loop has any metadata starting with Prefix. For example a 102 // Prefix of "llvm.loop.unroll." returns true if we have any unroll metadata. 103 static bool hasAnyUnrollPragma(const Loop *L, StringRef Prefix) { 104 if (MDNode *LoopID = L->getLoopID()) { 105 // First operand should refer to the loop id itself. 106 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 107 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 108 109 for (unsigned I = 1, E = LoopID->getNumOperands(); I < E; ++I) { 110 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(I)); 111 if (!MD) 112 continue; 113 114 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 115 if (!S) 116 continue; 117 118 if (S->getString().startswith(Prefix)) 119 return true; 120 } 121 } 122 return false; 123 } 124 125 // Returns true if the loop has an unroll_and_jam(enable) pragma. 126 static bool hasUnrollAndJamEnablePragma(const Loop *L) { 127 return getUnrollMetadataForLoop(L, "llvm.loop.unroll_and_jam.enable"); 128 } 129 130 // If loop has an unroll_and_jam_count pragma return the (necessarily 131 // positive) value from the pragma. Otherwise return 0. 132 static unsigned unrollAndJamCountPragmaValue(const Loop *L) { 133 MDNode *MD = getUnrollMetadataForLoop(L, "llvm.loop.unroll_and_jam.count"); 134 if (MD) { 135 assert(MD->getNumOperands() == 2 && 136 "Unroll count hint metadata should have two operands."); 137 unsigned Count = 138 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 139 assert(Count >= 1 && "Unroll count must be positive."); 140 return Count; 141 } 142 return 0; 143 } 144 145 // Returns loop size estimation for unrolled loop. 146 static uint64_t 147 getUnrollAndJammedLoopSize(unsigned LoopSize, 148 TargetTransformInfo::UnrollingPreferences &UP) { 149 assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!"); 150 return static_cast<uint64_t>(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns; 151 } 152 153 // Calculates unroll and jam count and writes it to UP.Count. Returns true if 154 // unroll count was set explicitly. 155 static bool computeUnrollAndJamCount( 156 Loop *L, Loop *SubLoop, const TargetTransformInfo &TTI, DominatorTree &DT, 157 LoopInfo *LI, ScalarEvolution &SE, 158 const SmallPtrSetImpl<const Value *> &EphValues, 159 OptimizationRemarkEmitter *ORE, unsigned OuterTripCount, 160 unsigned OuterTripMultiple, unsigned OuterLoopSize, unsigned InnerTripCount, 161 unsigned InnerLoopSize, TargetTransformInfo::UnrollingPreferences &UP, 162 TargetTransformInfo::PeelingPreferences &PP) { 163 // First up use computeUnrollCount from the loop unroller to get a count 164 // for unrolling the outer loop, plus any loops requiring explicit 165 // unrolling we leave to the unroller. This uses UP.Threshold / 166 // UP.PartialThreshold / UP.MaxCount to come up with sensible loop values. 167 // We have already checked that the loop has no unroll.* pragmas. 168 unsigned MaxTripCount = 0; 169 bool UseUpperBound = false; 170 bool ExplicitUnroll = computeUnrollCount( 171 L, TTI, DT, LI, SE, EphValues, ORE, OuterTripCount, MaxTripCount, 172 /*MaxOrZero*/ false, OuterTripMultiple, OuterLoopSize, UP, PP, 173 UseUpperBound); 174 if (ExplicitUnroll || UseUpperBound) { 175 // If the user explicitly set the loop as unrolled, dont UnJ it. Leave it 176 // for the unroller instead. 177 LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; explicit count set by " 178 "computeUnrollCount\n"); 179 UP.Count = 0; 180 return false; 181 } 182 183 // Override with any explicit Count from the "unroll-and-jam-count" option. 184 bool UserUnrollCount = UnrollAndJamCount.getNumOccurrences() > 0; 185 if (UserUnrollCount) { 186 UP.Count = UnrollAndJamCount; 187 UP.Force = true; 188 if (UP.AllowRemainder && 189 getUnrollAndJammedLoopSize(OuterLoopSize, UP) < UP.Threshold && 190 getUnrollAndJammedLoopSize(InnerLoopSize, UP) < 191 UP.UnrollAndJamInnerLoopThreshold) 192 return true; 193 } 194 195 // Check for unroll_and_jam pragmas 196 unsigned PragmaCount = unrollAndJamCountPragmaValue(L); 197 if (PragmaCount > 0) { 198 UP.Count = PragmaCount; 199 UP.Runtime = true; 200 UP.Force = true; 201 if ((UP.AllowRemainder || (OuterTripMultiple % PragmaCount == 0)) && 202 getUnrollAndJammedLoopSize(OuterLoopSize, UP) < UP.Threshold && 203 getUnrollAndJammedLoopSize(InnerLoopSize, UP) < 204 UP.UnrollAndJamInnerLoopThreshold) 205 return true; 206 } 207 208 bool PragmaEnableUnroll = hasUnrollAndJamEnablePragma(L); 209 bool ExplicitUnrollAndJamCount = PragmaCount > 0 || UserUnrollCount; 210 bool ExplicitUnrollAndJam = PragmaEnableUnroll || ExplicitUnrollAndJamCount; 211 212 // If the loop has an unrolling pragma, we want to be more aggressive with 213 // unrolling limits. 214 if (ExplicitUnrollAndJam) 215 UP.UnrollAndJamInnerLoopThreshold = PragmaUnrollAndJamThreshold; 216 217 if (!UP.AllowRemainder && getUnrollAndJammedLoopSize(InnerLoopSize, UP) >= 218 UP.UnrollAndJamInnerLoopThreshold) { 219 LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; can't create remainder and " 220 "inner loop too large\n"); 221 UP.Count = 0; 222 return false; 223 } 224 225 // We have a sensible limit for the outer loop, now adjust it for the inner 226 // loop and UP.UnrollAndJamInnerLoopThreshold. If the outer limit was set 227 // explicitly, we want to stick to it. 228 if (!ExplicitUnrollAndJamCount && UP.AllowRemainder) { 229 while (UP.Count != 0 && getUnrollAndJammedLoopSize(InnerLoopSize, UP) >= 230 UP.UnrollAndJamInnerLoopThreshold) 231 UP.Count--; 232 } 233 234 // If we are explicitly unroll and jamming, we are done. Otherwise there are a 235 // number of extra performance heuristics to check. 236 if (ExplicitUnrollAndJam) 237 return true; 238 239 // If the inner loop count is known and small, leave the entire loop nest to 240 // be the unroller 241 if (InnerTripCount && InnerLoopSize * InnerTripCount < UP.Threshold) { 242 LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; small inner loop count is " 243 "being left for the unroller\n"); 244 UP.Count = 0; 245 return false; 246 } 247 248 // Check for situations where UnJ is likely to be unprofitable. Including 249 // subloops with more than 1 block. 250 if (SubLoop->getBlocks().size() != 1) { 251 LLVM_DEBUG( 252 dbgs() << "Won't unroll-and-jam; More than one inner loop block\n"); 253 UP.Count = 0; 254 return false; 255 } 256 257 // Limit to loops where there is something to gain from unrolling and 258 // jamming the loop. In this case, look for loads that are invariant in the 259 // outer loop and can become shared. 260 unsigned NumInvariant = 0; 261 for (BasicBlock *BB : SubLoop->getBlocks()) { 262 for (Instruction &I : *BB) { 263 if (auto *Ld = dyn_cast<LoadInst>(&I)) { 264 Value *V = Ld->getPointerOperand(); 265 const SCEV *LSCEV = SE.getSCEVAtScope(V, L); 266 if (SE.isLoopInvariant(LSCEV, L)) 267 NumInvariant++; 268 } 269 } 270 } 271 if (NumInvariant == 0) { 272 LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; No loop invariant loads\n"); 273 UP.Count = 0; 274 return false; 275 } 276 277 return false; 278 } 279 280 static LoopUnrollResult 281 tryToUnrollAndJamLoop(Loop *L, DominatorTree &DT, LoopInfo *LI, 282 ScalarEvolution &SE, const TargetTransformInfo &TTI, 283 AssumptionCache &AC, DependenceInfo &DI, 284 OptimizationRemarkEmitter &ORE, int OptLevel) { 285 TargetTransformInfo::UnrollingPreferences UP = 286 gatherUnrollingPreferences(L, SE, TTI, nullptr, nullptr, OptLevel, None, 287 None, None, None, None, None); 288 TargetTransformInfo::PeelingPreferences PP = 289 gatherPeelingPreferences(L, SE, TTI, None, None); 290 if (AllowUnrollAndJam.getNumOccurrences() > 0) 291 UP.UnrollAndJam = AllowUnrollAndJam; 292 if (UnrollAndJamThreshold.getNumOccurrences() > 0) 293 UP.UnrollAndJamInnerLoopThreshold = UnrollAndJamThreshold; 294 // Exit early if unrolling is disabled. 295 if (!UP.UnrollAndJam || UP.UnrollAndJamInnerLoopThreshold == 0) 296 return LoopUnrollResult::Unmodified; 297 298 LLVM_DEBUG(dbgs() << "Loop Unroll and Jam: F[" 299 << L->getHeader()->getParent()->getName() << "] Loop %" 300 << L->getHeader()->getName() << "\n"); 301 302 TransformationMode EnableMode = hasUnrollAndJamTransformation(L); 303 if (EnableMode & TM_Disable) 304 return LoopUnrollResult::Unmodified; 305 306 // A loop with any unroll pragma (enabling/disabling/count/etc) is left for 307 // the unroller, so long as it does not explicitly have unroll_and_jam 308 // metadata. This means #pragma nounroll will disable unroll and jam as well 309 // as unrolling 310 if (hasAnyUnrollPragma(L, "llvm.loop.unroll.") && 311 !hasAnyUnrollPragma(L, "llvm.loop.unroll_and_jam.")) { 312 LLVM_DEBUG(dbgs() << " Disabled due to pragma.\n"); 313 return LoopUnrollResult::Unmodified; 314 } 315 316 if (!isSafeToUnrollAndJam(L, SE, DT, DI, *LI)) { 317 LLVM_DEBUG(dbgs() << " Disabled due to not being safe.\n"); 318 return LoopUnrollResult::Unmodified; 319 } 320 321 // Approximate the loop size and collect useful info 322 unsigned NumInlineCandidates; 323 bool NotDuplicatable; 324 bool Convergent; 325 SmallPtrSet<const Value *, 32> EphValues; 326 CodeMetrics::collectEphemeralValues(L, &AC, EphValues); 327 Loop *SubLoop = L->getSubLoops()[0]; 328 unsigned InnerLoopSize = 329 ApproximateLoopSize(SubLoop, NumInlineCandidates, NotDuplicatable, 330 Convergent, TTI, EphValues, UP.BEInsns); 331 unsigned OuterLoopSize = 332 ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent, 333 TTI, EphValues, UP.BEInsns); 334 LLVM_DEBUG(dbgs() << " Outer Loop Size: " << OuterLoopSize << "\n"); 335 LLVM_DEBUG(dbgs() << " Inner Loop Size: " << InnerLoopSize << "\n"); 336 if (NotDuplicatable) { 337 LLVM_DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable " 338 "instructions.\n"); 339 return LoopUnrollResult::Unmodified; 340 } 341 if (NumInlineCandidates != 0) { 342 LLVM_DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 343 return LoopUnrollResult::Unmodified; 344 } 345 if (Convergent) { 346 LLVM_DEBUG( 347 dbgs() << " Not unrolling loop with convergent instructions.\n"); 348 return LoopUnrollResult::Unmodified; 349 } 350 351 // Save original loop IDs for after the transformation. 352 MDNode *OrigOuterLoopID = L->getLoopID(); 353 MDNode *OrigSubLoopID = SubLoop->getLoopID(); 354 355 // To assign the loop id of the epilogue, assign it before unrolling it so it 356 // is applied to every inner loop of the epilogue. We later apply the loop ID 357 // for the jammed inner loop. 358 Optional<MDNode *> NewInnerEpilogueLoopID = makeFollowupLoopID( 359 OrigOuterLoopID, {LLVMLoopUnrollAndJamFollowupAll, 360 LLVMLoopUnrollAndJamFollowupRemainderInner}); 361 if (NewInnerEpilogueLoopID.hasValue()) 362 SubLoop->setLoopID(NewInnerEpilogueLoopID.getValue()); 363 364 // Find trip count and trip multiple 365 BasicBlock *Latch = L->getLoopLatch(); 366 BasicBlock *SubLoopLatch = SubLoop->getLoopLatch(); 367 unsigned OuterTripCount = SE.getSmallConstantTripCount(L, Latch); 368 unsigned OuterTripMultiple = SE.getSmallConstantTripMultiple(L, Latch); 369 unsigned InnerTripCount = SE.getSmallConstantTripCount(SubLoop, SubLoopLatch); 370 371 // Decide if, and by how much, to unroll 372 bool IsCountSetExplicitly = computeUnrollAndJamCount( 373 L, SubLoop, TTI, DT, LI, SE, EphValues, &ORE, OuterTripCount, 374 OuterTripMultiple, OuterLoopSize, InnerTripCount, InnerLoopSize, UP, PP); 375 if (UP.Count <= 1) 376 return LoopUnrollResult::Unmodified; 377 // Unroll factor (Count) must be less or equal to TripCount. 378 if (OuterTripCount && UP.Count > OuterTripCount) 379 UP.Count = OuterTripCount; 380 381 Loop *EpilogueOuterLoop = nullptr; 382 LoopUnrollResult UnrollResult = UnrollAndJamLoop( 383 L, UP.Count, OuterTripCount, OuterTripMultiple, UP.UnrollRemainder, LI, 384 &SE, &DT, &AC, &TTI, &ORE, &EpilogueOuterLoop); 385 386 // Assign new loop attributes. 387 if (EpilogueOuterLoop) { 388 Optional<MDNode *> NewOuterEpilogueLoopID = makeFollowupLoopID( 389 OrigOuterLoopID, {LLVMLoopUnrollAndJamFollowupAll, 390 LLVMLoopUnrollAndJamFollowupRemainderOuter}); 391 if (NewOuterEpilogueLoopID.hasValue()) 392 EpilogueOuterLoop->setLoopID(NewOuterEpilogueLoopID.getValue()); 393 } 394 395 Optional<MDNode *> NewInnerLoopID = 396 makeFollowupLoopID(OrigOuterLoopID, {LLVMLoopUnrollAndJamFollowupAll, 397 LLVMLoopUnrollAndJamFollowupInner}); 398 if (NewInnerLoopID.hasValue()) 399 SubLoop->setLoopID(NewInnerLoopID.getValue()); 400 else 401 SubLoop->setLoopID(OrigSubLoopID); 402 403 if (UnrollResult == LoopUnrollResult::PartiallyUnrolled) { 404 Optional<MDNode *> NewOuterLoopID = makeFollowupLoopID( 405 OrigOuterLoopID, 406 {LLVMLoopUnrollAndJamFollowupAll, LLVMLoopUnrollAndJamFollowupOuter}); 407 if (NewOuterLoopID.hasValue()) { 408 L->setLoopID(NewOuterLoopID.getValue()); 409 410 // Do not setLoopAlreadyUnrolled if a followup was given. 411 return UnrollResult; 412 } 413 } 414 415 // If loop has an unroll count pragma or unrolled by explicitly set count 416 // mark loop as unrolled to prevent unrolling beyond that requested. 417 if (UnrollResult != LoopUnrollResult::FullyUnrolled && IsCountSetExplicitly) 418 L->setLoopAlreadyUnrolled(); 419 420 return UnrollResult; 421 } 422 423 static bool tryToUnrollAndJamLoop(Function &F, DominatorTree &DT, LoopInfo &LI, 424 ScalarEvolution &SE, 425 const TargetTransformInfo &TTI, 426 AssumptionCache &AC, DependenceInfo &DI, 427 OptimizationRemarkEmitter &ORE, 428 int OptLevel) { 429 bool DidSomething = false; 430 431 // The loop unroll and jam pass requires loops to be in simplified form, and 432 // also needs LCSSA. Since simplification may add new inner loops, it has to 433 // run before the legality and profitability checks. This means running the 434 // loop unroll and jam pass will simplify all loops, regardless of whether 435 // anything end up being unroll and jammed. 436 for (auto &L : LI) { 437 DidSomething |= 438 simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */); 439 DidSomething |= formLCSSARecursively(*L, DT, &LI, &SE); 440 } 441 442 // Add the loop nests in the reverse order of LoopInfo. See method 443 // declaration. 444 SmallPriorityWorklist<Loop *, 4> Worklist; 445 appendLoopsToWorklist(LI, Worklist); 446 while (!Worklist.empty()) { 447 Loop *L = Worklist.pop_back_val(); 448 LoopUnrollResult Result = 449 tryToUnrollAndJamLoop(L, DT, &LI, SE, TTI, AC, DI, ORE, OptLevel); 450 if (Result != LoopUnrollResult::Unmodified) 451 DidSomething = true; 452 } 453 454 return DidSomething; 455 } 456 457 namespace { 458 459 class LoopUnrollAndJam : public FunctionPass { 460 public: 461 static char ID; // Pass ID, replacement for typeid 462 unsigned OptLevel; 463 464 LoopUnrollAndJam(int OptLevel = 2) : FunctionPass(ID), OptLevel(OptLevel) { 465 initializeLoopUnrollAndJamPass(*PassRegistry::getPassRegistry()); 466 } 467 468 bool runOnFunction(Function &F) override { 469 if (skipFunction(F)) 470 return false; 471 472 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 473 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 474 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 475 const TargetTransformInfo &TTI = 476 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 477 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 478 auto &DI = getAnalysis<DependenceAnalysisWrapperPass>().getDI(); 479 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); 480 481 return tryToUnrollAndJamLoop(F, DT, LI, SE, TTI, AC, DI, ORE, OptLevel); 482 } 483 484 /// This transformation requires natural loop information & requires that 485 /// loop preheaders be inserted into the CFG... 486 void getAnalysisUsage(AnalysisUsage &AU) const override { 487 AU.addRequired<DominatorTreeWrapperPass>(); 488 AU.addRequired<LoopInfoWrapperPass>(); 489 AU.addRequired<ScalarEvolutionWrapperPass>(); 490 AU.addRequired<TargetTransformInfoWrapperPass>(); 491 AU.addRequired<AssumptionCacheTracker>(); 492 AU.addRequired<DependenceAnalysisWrapperPass>(); 493 AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); 494 } 495 }; 496 497 } // end anonymous namespace 498 499 char LoopUnrollAndJam::ID = 0; 500 501 INITIALIZE_PASS_BEGIN(LoopUnrollAndJam, "loop-unroll-and-jam", 502 "Unroll and Jam loops", false, false) 503 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 504 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 505 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 506 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 507 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 508 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass) 509 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) 510 INITIALIZE_PASS_END(LoopUnrollAndJam, "loop-unroll-and-jam", 511 "Unroll and Jam loops", false, false) 512 513 Pass *llvm::createLoopUnrollAndJamPass(int OptLevel) { 514 return new LoopUnrollAndJam(OptLevel); 515 } 516 517 PreservedAnalyses LoopUnrollAndJamPass::run(Function &F, 518 FunctionAnalysisManager &AM) { 519 ScalarEvolution &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 520 LoopInfo &LI = AM.getResult<LoopAnalysis>(F); 521 TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F); 522 AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(F); 523 DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F); 524 DependenceInfo &DI = AM.getResult<DependenceAnalysis>(F); 525 OptimizationRemarkEmitter &ORE = 526 AM.getResult<OptimizationRemarkEmitterAnalysis>(F); 527 528 if (!tryToUnrollAndJamLoop(F, DT, LI, SE, TTI, AC, DI, ORE, OptLevel)) 529 return PreservedAnalyses::all(); 530 531 return getLoopPassPreservedAnalyses(); 532 } 533