1 //===- LoopPeel.cpp -------------------------------------------------------===// 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 // Loop Peeling Utilities. 10 //===----------------------------------------------------------------------===// 11 12 #include "llvm/Transforms/Utils/LoopPeel.h" 13 #include "llvm/ADT/DenseMap.h" 14 #include "llvm/ADT/Optional.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/Statistic.h" 17 #include "llvm/Analysis/Loads.h" 18 #include "llvm/Analysis/LoopInfo.h" 19 #include "llvm/Analysis/LoopIterator.h" 20 #include "llvm/Analysis/ScalarEvolution.h" 21 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 22 #include "llvm/Analysis/TargetTransformInfo.h" 23 #include "llvm/IR/BasicBlock.h" 24 #include "llvm/IR/Dominators.h" 25 #include "llvm/IR/Function.h" 26 #include "llvm/IR/InstrTypes.h" 27 #include "llvm/IR/Instruction.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/LLVMContext.h" 30 #include "llvm/IR/MDBuilder.h" 31 #include "llvm/IR/PatternMatch.h" 32 #include "llvm/Support/Casting.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Support/raw_ostream.h" 36 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 37 #include "llvm/Transforms/Utils/Cloning.h" 38 #include "llvm/Transforms/Utils/LoopSimplify.h" 39 #include "llvm/Transforms/Utils/LoopUtils.h" 40 #include "llvm/Transforms/Utils/ValueMapper.h" 41 #include <algorithm> 42 #include <cassert> 43 #include <cstdint> 44 45 using namespace llvm; 46 using namespace llvm::PatternMatch; 47 48 #define DEBUG_TYPE "loop-peel" 49 50 STATISTIC(NumPeeled, "Number of loops peeled"); 51 52 static cl::opt<unsigned> UnrollPeelCount( 53 "unroll-peel-count", cl::Hidden, 54 cl::desc("Set the unroll peeling count, for testing purposes")); 55 56 static cl::opt<bool> 57 UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden, 58 cl::desc("Allows loops to be peeled when the dynamic " 59 "trip count is known to be low.")); 60 61 static cl::opt<bool> 62 UnrollAllowLoopNestsPeeling("unroll-allow-loop-nests-peeling", 63 cl::init(false), cl::Hidden, 64 cl::desc("Allows loop nests to be peeled.")); 65 66 static cl::opt<unsigned> UnrollPeelMaxCount( 67 "unroll-peel-max-count", cl::init(7), cl::Hidden, 68 cl::desc("Max average trip count which will cause loop peeling.")); 69 70 static cl::opt<unsigned> UnrollForcePeelCount( 71 "unroll-force-peel-count", cl::init(0), cl::Hidden, 72 cl::desc("Force a peel count regardless of profiling information.")); 73 74 static const char *PeeledCountMetaData = "llvm.loop.peeled.count"; 75 76 // Check whether we are capable of peeling this loop. 77 bool llvm::canPeel(Loop *L) { 78 // Make sure the loop is in simplified form 79 if (!L->isLoopSimplifyForm()) 80 return false; 81 82 // Don't try to peel loops where the latch is not the exiting block. 83 // This can be an indication of two different things: 84 // 1) The loop is not rotated. 85 // 2) The loop contains irreducible control flow that involves the latch. 86 const BasicBlock *Latch = L->getLoopLatch(); 87 if (!L->isLoopExiting(Latch)) 88 return false; 89 90 // Peeling is only supported if the latch is a branch. 91 if (!isa<BranchInst>(Latch->getTerminator())) 92 return false; 93 94 SmallVector<BasicBlock *, 4> Exits; 95 L->getUniqueNonLatchExitBlocks(Exits); 96 // The latch must either be the only exiting block or all non-latch exit 97 // blocks have either a deopt or unreachable terminator or compose a chain of 98 // blocks where the last one is either deopt or unreachable terminated. Both 99 // deopt and unreachable terminators are a strong indication they are not 100 // taken. Note that this is a profitability check, not a legality check. Also 101 // note that LoopPeeling currently can only update the branch weights of latch 102 // blocks and branch weights to blocks with deopt or unreachable do not need 103 // updating. 104 return llvm::all_of(Exits, IsBlockFollowedByDeoptOrUnreachable); 105 } 106 107 // This function calculates the number of iterations after which the given Phi 108 // becomes an invariant. The pre-calculated values are memorized in the map. The 109 // function (shortcut is I) is calculated according to the following definition: 110 // Given %x = phi <Inputs from above the loop>, ..., [%y, %back.edge]. 111 // If %y is a loop invariant, then I(%x) = 1. 112 // If %y is a Phi from the loop header, I(%x) = I(%y) + 1. 113 // Otherwise, I(%x) is infinite. 114 // TODO: Actually if %y is an expression that depends only on Phi %z and some 115 // loop invariants, we can estimate I(%x) = I(%z) + 1. The example 116 // looks like: 117 // %x = phi(0, %a), <-- becomes invariant starting from 3rd iteration. 118 // %y = phi(0, 5), 119 // %a = %y + 1. 120 static Optional<unsigned> calculateIterationsToInvariance( 121 PHINode *Phi, Loop *L, BasicBlock *BackEdge, 122 SmallDenseMap<PHINode *, Optional<unsigned> > &IterationsToInvariance) { 123 assert(Phi->getParent() == L->getHeader() && 124 "Non-loop Phi should not be checked for turning into invariant."); 125 assert(BackEdge == L->getLoopLatch() && "Wrong latch?"); 126 // If we already know the answer, take it from the map. 127 auto I = IterationsToInvariance.find(Phi); 128 if (I != IterationsToInvariance.end()) 129 return I->second; 130 131 // Otherwise we need to analyze the input from the back edge. 132 Value *Input = Phi->getIncomingValueForBlock(BackEdge); 133 // Place infinity to map to avoid infinite recursion for cycled Phis. Such 134 // cycles can never stop on an invariant. 135 IterationsToInvariance[Phi] = None; 136 Optional<unsigned> ToInvariance = None; 137 138 if (L->isLoopInvariant(Input)) 139 ToInvariance = 1u; 140 else if (PHINode *IncPhi = dyn_cast<PHINode>(Input)) { 141 // Only consider Phis in header block. 142 if (IncPhi->getParent() != L->getHeader()) 143 return None; 144 // If the input becomes an invariant after X iterations, then our Phi 145 // becomes an invariant after X + 1 iterations. 146 auto InputToInvariance = calculateIterationsToInvariance( 147 IncPhi, L, BackEdge, IterationsToInvariance); 148 if (InputToInvariance) 149 ToInvariance = *InputToInvariance + 1u; 150 } 151 152 // If we found that this Phi lies in an invariant chain, update the map. 153 if (ToInvariance) 154 IterationsToInvariance[Phi] = ToInvariance; 155 return ToInvariance; 156 } 157 158 // Try to find any invariant memory reads that will become dereferenceable in 159 // the remainder loop after peeling. The load must also be used (transitively) 160 // by an exit condition. Returns the number of iterations to peel off (at the 161 // moment either 0 or 1). 162 static unsigned peelToTurnInvariantLoadsDerefencebale(Loop &L, 163 DominatorTree &DT) { 164 // Skip loops with a single exiting block, because there should be no benefit 165 // for the heuristic below. 166 if (L.getExitingBlock()) 167 return 0; 168 169 // All non-latch exit blocks must have an UnreachableInst terminator. 170 // Otherwise the heuristic below may not be profitable. 171 SmallVector<BasicBlock *, 4> Exits; 172 L.getUniqueNonLatchExitBlocks(Exits); 173 if (any_of(Exits, [](const BasicBlock *BB) { 174 return !isa<UnreachableInst>(BB->getTerminator()); 175 })) 176 return 0; 177 178 // Now look for invariant loads that dominate the latch and are not known to 179 // be dereferenceable. If there are such loads and no writes, they will become 180 // dereferenceable in the loop if the first iteration is peeled off. Also 181 // collect the set of instructions controlled by such loads. Only peel if an 182 // exit condition uses (transitively) such a load. 183 BasicBlock *Header = L.getHeader(); 184 BasicBlock *Latch = L.getLoopLatch(); 185 SmallPtrSet<Value *, 8> LoadUsers; 186 const DataLayout &DL = L.getHeader()->getModule()->getDataLayout(); 187 for (BasicBlock *BB : L.blocks()) { 188 for (Instruction &I : *BB) { 189 if (I.mayWriteToMemory()) 190 return 0; 191 192 auto Iter = LoadUsers.find(&I); 193 if (Iter != LoadUsers.end()) { 194 for (Value *U : I.users()) 195 LoadUsers.insert(U); 196 } 197 // Do not look for reads in the header; they can already be hoisted 198 // without peeling. 199 if (BB == Header) 200 continue; 201 if (auto *LI = dyn_cast<LoadInst>(&I)) { 202 Value *Ptr = LI->getPointerOperand(); 203 if (DT.dominates(BB, Latch) && L.isLoopInvariant(Ptr) && 204 !isDereferenceablePointer(Ptr, LI->getType(), DL, LI, &DT)) 205 for (Value *U : I.users()) 206 LoadUsers.insert(U); 207 } 208 } 209 } 210 SmallVector<BasicBlock *> ExitingBlocks; 211 L.getExitingBlocks(ExitingBlocks); 212 if (any_of(ExitingBlocks, [&LoadUsers](BasicBlock *Exiting) { 213 return LoadUsers.contains(Exiting->getTerminator()); 214 })) 215 return 1; 216 return 0; 217 } 218 219 // Return the number of iterations to peel off that make conditions in the 220 // body true/false. For example, if we peel 2 iterations off the loop below, 221 // the condition i < 2 can be evaluated at compile time. 222 // for (i = 0; i < n; i++) 223 // if (i < 2) 224 // .. 225 // else 226 // .. 227 // } 228 static unsigned countToEliminateCompares(Loop &L, unsigned MaxPeelCount, 229 ScalarEvolution &SE) { 230 assert(L.isLoopSimplifyForm() && "Loop needs to be in loop simplify form"); 231 unsigned DesiredPeelCount = 0; 232 233 for (auto *BB : L.blocks()) { 234 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 235 if (!BI || BI->isUnconditional()) 236 continue; 237 238 // Ignore loop exit condition. 239 if (L.getLoopLatch() == BB) 240 continue; 241 242 Value *Condition = BI->getCondition(); 243 Value *LeftVal, *RightVal; 244 CmpInst::Predicate Pred; 245 if (!match(Condition, m_ICmp(Pred, m_Value(LeftVal), m_Value(RightVal)))) 246 continue; 247 248 const SCEV *LeftSCEV = SE.getSCEV(LeftVal); 249 const SCEV *RightSCEV = SE.getSCEV(RightVal); 250 251 // Do not consider predicates that are known to be true or false 252 // independently of the loop iteration. 253 if (SE.evaluatePredicate(Pred, LeftSCEV, RightSCEV)) 254 continue; 255 256 // Check if we have a condition with one AddRec and one non AddRec 257 // expression. Normalize LeftSCEV to be the AddRec. 258 if (!isa<SCEVAddRecExpr>(LeftSCEV)) { 259 if (isa<SCEVAddRecExpr>(RightSCEV)) { 260 std::swap(LeftSCEV, RightSCEV); 261 Pred = ICmpInst::getSwappedPredicate(Pred); 262 } else 263 continue; 264 } 265 266 const SCEVAddRecExpr *LeftAR = cast<SCEVAddRecExpr>(LeftSCEV); 267 268 // Avoid huge SCEV computations in the loop below, make sure we only 269 // consider AddRecs of the loop we are trying to peel. 270 if (!LeftAR->isAffine() || LeftAR->getLoop() != &L) 271 continue; 272 if (!(ICmpInst::isEquality(Pred) && LeftAR->hasNoSelfWrap()) && 273 !SE.getMonotonicPredicateType(LeftAR, Pred)) 274 continue; 275 276 // Check if extending the current DesiredPeelCount lets us evaluate Pred 277 // or !Pred in the loop body statically. 278 unsigned NewPeelCount = DesiredPeelCount; 279 280 const SCEV *IterVal = LeftAR->evaluateAtIteration( 281 SE.getConstant(LeftSCEV->getType(), NewPeelCount), SE); 282 283 // If the original condition is not known, get the negated predicate 284 // (which holds on the else branch) and check if it is known. This allows 285 // us to peel of iterations that make the original condition false. 286 if (!SE.isKnownPredicate(Pred, IterVal, RightSCEV)) 287 Pred = ICmpInst::getInversePredicate(Pred); 288 289 const SCEV *Step = LeftAR->getStepRecurrence(SE); 290 const SCEV *NextIterVal = SE.getAddExpr(IterVal, Step); 291 auto PeelOneMoreIteration = [&IterVal, &NextIterVal, &SE, Step, 292 &NewPeelCount]() { 293 IterVal = NextIterVal; 294 NextIterVal = SE.getAddExpr(IterVal, Step); 295 NewPeelCount++; 296 }; 297 298 auto CanPeelOneMoreIteration = [&NewPeelCount, &MaxPeelCount]() { 299 return NewPeelCount < MaxPeelCount; 300 }; 301 302 while (CanPeelOneMoreIteration() && 303 SE.isKnownPredicate(Pred, IterVal, RightSCEV)) 304 PeelOneMoreIteration(); 305 306 // With *that* peel count, does the predicate !Pred become known in the 307 // first iteration of the loop body after peeling? 308 if (!SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), IterVal, 309 RightSCEV)) 310 continue; // If not, give up. 311 312 // However, for equality comparisons, that isn't always sufficient to 313 // eliminate the comparsion in loop body, we may need to peel one more 314 // iteration. See if that makes !Pred become unknown again. 315 if (ICmpInst::isEquality(Pred) && 316 !SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), NextIterVal, 317 RightSCEV) && 318 !SE.isKnownPredicate(Pred, IterVal, RightSCEV) && 319 SE.isKnownPredicate(Pred, NextIterVal, RightSCEV)) { 320 if (!CanPeelOneMoreIteration()) 321 continue; // Need to peel one more iteration, but can't. Give up. 322 PeelOneMoreIteration(); // Great! 323 } 324 325 DesiredPeelCount = std::max(DesiredPeelCount, NewPeelCount); 326 } 327 328 return DesiredPeelCount; 329 } 330 331 /// This "heuristic" exactly matches implicit behavior which used to exist 332 /// inside getLoopEstimatedTripCount. It was added here to keep an 333 /// improvement inside that API from causing peeling to become more agressive. 334 /// This should probably be removed. 335 static bool violatesLegacyMultiExitLoopCheck(Loop *L) { 336 BasicBlock *Latch = L->getLoopLatch(); 337 if (!Latch) 338 return true; 339 340 BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator()); 341 if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch)) 342 return true; 343 344 assert((LatchBR->getSuccessor(0) == L->getHeader() || 345 LatchBR->getSuccessor(1) == L->getHeader()) && 346 "At least one edge out of the latch must go to the header"); 347 348 SmallVector<BasicBlock *, 4> ExitBlocks; 349 L->getUniqueNonLatchExitBlocks(ExitBlocks); 350 return any_of(ExitBlocks, [](const BasicBlock *EB) { 351 return !EB->getTerminatingDeoptimizeCall(); 352 }); 353 } 354 355 356 // Return the number of iterations we want to peel off. 357 void llvm::computePeelCount(Loop *L, unsigned LoopSize, 358 TargetTransformInfo::PeelingPreferences &PP, 359 unsigned TripCount, DominatorTree &DT, 360 ScalarEvolution &SE, unsigned Threshold) { 361 assert(LoopSize > 0 && "Zero loop size is not allowed!"); 362 // Save the PP.PeelCount value set by the target in 363 // TTI.getPeelingPreferences or by the flag -unroll-peel-count. 364 unsigned TargetPeelCount = PP.PeelCount; 365 PP.PeelCount = 0; 366 if (!canPeel(L)) 367 return; 368 369 // Only try to peel innermost loops by default. 370 // The constraint can be relaxed by the target in TTI.getPeelingPreferences 371 // or by the flag -unroll-allow-loop-nests-peeling. 372 if (!PP.AllowLoopNestsPeeling && !L->isInnermost()) 373 return; 374 375 // If the user provided a peel count, use that. 376 bool UserPeelCount = UnrollForcePeelCount.getNumOccurrences() > 0; 377 if (UserPeelCount) { 378 LLVM_DEBUG(dbgs() << "Force-peeling first " << UnrollForcePeelCount 379 << " iterations.\n"); 380 PP.PeelCount = UnrollForcePeelCount; 381 PP.PeelProfiledIterations = true; 382 return; 383 } 384 385 // Skip peeling if it's disabled. 386 if (!PP.AllowPeeling) 387 return; 388 389 // Check that we can peel at least one iteration. 390 if (2 * LoopSize > Threshold) 391 return; 392 393 unsigned AlreadyPeeled = 0; 394 if (auto Peeled = getOptionalIntLoopAttribute(L, PeeledCountMetaData)) 395 AlreadyPeeled = *Peeled; 396 // Stop if we already peeled off the maximum number of iterations. 397 if (AlreadyPeeled >= UnrollPeelMaxCount) 398 return; 399 400 // Here we try to get rid of Phis which become invariants after 1, 2, ..., N 401 // iterations of the loop. For this we compute the number for iterations after 402 // which every Phi is guaranteed to become an invariant, and try to peel the 403 // maximum number of iterations among these values, thus turning all those 404 // Phis into invariants. 405 406 // Store the pre-calculated values here. 407 SmallDenseMap<PHINode *, Optional<unsigned>> IterationsToInvariance; 408 // Now go through all Phis to calculate their the number of iterations they 409 // need to become invariants. 410 // Start the max computation with the PP.PeelCount value set by the target 411 // in TTI.getPeelingPreferences or by the flag -unroll-peel-count. 412 unsigned DesiredPeelCount = TargetPeelCount; 413 BasicBlock *BackEdge = L->getLoopLatch(); 414 assert(BackEdge && "Loop is not in simplified form?"); 415 for (auto BI = L->getHeader()->begin(); isa<PHINode>(&*BI); ++BI) { 416 PHINode *Phi = cast<PHINode>(&*BI); 417 auto ToInvariance = calculateIterationsToInvariance(Phi, L, BackEdge, 418 IterationsToInvariance); 419 if (ToInvariance) 420 DesiredPeelCount = std::max(DesiredPeelCount, *ToInvariance); 421 } 422 423 // Pay respect to limitations implied by loop size and the max peel count. 424 unsigned MaxPeelCount = UnrollPeelMaxCount; 425 MaxPeelCount = std::min(MaxPeelCount, Threshold / LoopSize - 1); 426 427 DesiredPeelCount = std::max(DesiredPeelCount, 428 countToEliminateCompares(*L, MaxPeelCount, SE)); 429 430 if (DesiredPeelCount == 0) 431 DesiredPeelCount = peelToTurnInvariantLoadsDerefencebale(*L, DT); 432 433 if (DesiredPeelCount > 0) { 434 DesiredPeelCount = std::min(DesiredPeelCount, MaxPeelCount); 435 // Consider max peel count limitation. 436 assert(DesiredPeelCount > 0 && "Wrong loop size estimation?"); 437 if (DesiredPeelCount + AlreadyPeeled <= UnrollPeelMaxCount) { 438 LLVM_DEBUG(dbgs() << "Peel " << DesiredPeelCount 439 << " iteration(s) to turn" 440 << " some Phis into invariants.\n"); 441 PP.PeelCount = DesiredPeelCount; 442 PP.PeelProfiledIterations = false; 443 return; 444 } 445 } 446 447 // Bail if we know the statically calculated trip count. 448 // In this case we rather prefer partial unrolling. 449 if (TripCount) 450 return; 451 452 // Do not apply profile base peeling if it is disabled. 453 if (!PP.PeelProfiledIterations) 454 return; 455 // If we don't know the trip count, but have reason to believe the average 456 // trip count is low, peeling should be beneficial, since we will usually 457 // hit the peeled section. 458 // We only do this in the presence of profile information, since otherwise 459 // our estimates of the trip count are not reliable enough. 460 if (L->getHeader()->getParent()->hasProfileData()) { 461 if (violatesLegacyMultiExitLoopCheck(L)) 462 return; 463 Optional<unsigned> EstimatedTripCount = getLoopEstimatedTripCount(L); 464 if (!EstimatedTripCount) 465 return; 466 467 LLVM_DEBUG(dbgs() << "Profile-based estimated trip count is " 468 << *EstimatedTripCount << "\n"); 469 470 if (*EstimatedTripCount) { 471 if (*EstimatedTripCount + AlreadyPeeled <= MaxPeelCount) { 472 unsigned PeelCount = *EstimatedTripCount; 473 LLVM_DEBUG(dbgs() << "Peeling first " << PeelCount << " iterations.\n"); 474 PP.PeelCount = PeelCount; 475 return; 476 } 477 LLVM_DEBUG(dbgs() << "Already peel count: " << AlreadyPeeled << "\n"); 478 LLVM_DEBUG(dbgs() << "Max peel count: " << UnrollPeelMaxCount << "\n"); 479 LLVM_DEBUG(dbgs() << "Loop cost: " << LoopSize << "\n"); 480 LLVM_DEBUG(dbgs() << "Max peel cost: " << Threshold << "\n"); 481 LLVM_DEBUG(dbgs() << "Max peel count by cost: " 482 << (Threshold / LoopSize - 1) << "\n"); 483 } 484 } 485 } 486 487 /// Update the branch weights of the latch of a peeled-off loop 488 /// iteration. 489 /// This sets the branch weights for the latch of the recently peeled off loop 490 /// iteration correctly. 491 /// Let F is a weight of the edge from latch to header. 492 /// Let E is a weight of the edge from latch to exit. 493 /// F/(F+E) is a probability to go to loop and E/(F+E) is a probability to 494 /// go to exit. 495 /// Then, Estimated TripCount = F / E. 496 /// For I-th (counting from 0) peeled off iteration we set the the weights for 497 /// the peeled latch as (TC - I, 1). It gives us reasonable distribution, 498 /// The probability to go to exit 1/(TC-I) increases. At the same time 499 /// the estimated trip count of remaining loop reduces by I. 500 /// To avoid dealing with division rounding we can just multiple both part 501 /// of weights to E and use weight as (F - I * E, E). 502 /// 503 /// \param Header The copy of the header block that belongs to next iteration. 504 /// \param LatchBR The copy of the latch branch that belongs to this iteration. 505 /// \param[in,out] FallThroughWeight The weight of the edge from latch to 506 /// header before peeling (in) and after peeled off one iteration (out). 507 static void updateBranchWeights(BasicBlock *Header, BranchInst *LatchBR, 508 uint64_t ExitWeight, 509 uint64_t &FallThroughWeight) { 510 // FallThroughWeight is 0 means that there is no branch weights on original 511 // latch block or estimated trip count is zero. 512 if (!FallThroughWeight) 513 return; 514 515 unsigned HeaderIdx = (LatchBR->getSuccessor(0) == Header ? 0 : 1); 516 MDBuilder MDB(LatchBR->getContext()); 517 MDNode *WeightNode = 518 HeaderIdx ? MDB.createBranchWeights(ExitWeight, FallThroughWeight) 519 : MDB.createBranchWeights(FallThroughWeight, ExitWeight); 520 LatchBR->setMetadata(LLVMContext::MD_prof, WeightNode); 521 FallThroughWeight = 522 FallThroughWeight > ExitWeight ? FallThroughWeight - ExitWeight : 1; 523 } 524 525 /// Initialize the weights. 526 /// 527 /// \param Header The header block. 528 /// \param LatchBR The latch branch. 529 /// \param[out] ExitWeight The weight of the edge from Latch to Exit. 530 /// \param[out] FallThroughWeight The weight of the edge from Latch to Header. 531 static void initBranchWeights(BasicBlock *Header, BranchInst *LatchBR, 532 uint64_t &ExitWeight, 533 uint64_t &FallThroughWeight) { 534 uint64_t TrueWeight, FalseWeight; 535 if (!LatchBR->extractProfMetadata(TrueWeight, FalseWeight)) 536 return; 537 unsigned HeaderIdx = LatchBR->getSuccessor(0) == Header ? 0 : 1; 538 ExitWeight = HeaderIdx ? TrueWeight : FalseWeight; 539 FallThroughWeight = HeaderIdx ? FalseWeight : TrueWeight; 540 } 541 542 /// Update the weights of original Latch block after peeling off all iterations. 543 /// 544 /// \param Header The header block. 545 /// \param LatchBR The latch branch. 546 /// \param ExitWeight The weight of the edge from Latch to Exit. 547 /// \param FallThroughWeight The weight of the edge from Latch to Header. 548 static void fixupBranchWeights(BasicBlock *Header, BranchInst *LatchBR, 549 uint64_t ExitWeight, 550 uint64_t FallThroughWeight) { 551 // FallThroughWeight is 0 means that there is no branch weights on original 552 // latch block or estimated trip count is zero. 553 if (!FallThroughWeight) 554 return; 555 556 // Sets the branch weights on the loop exit. 557 MDBuilder MDB(LatchBR->getContext()); 558 unsigned HeaderIdx = LatchBR->getSuccessor(0) == Header ? 0 : 1; 559 MDNode *WeightNode = 560 HeaderIdx ? MDB.createBranchWeights(ExitWeight, FallThroughWeight) 561 : MDB.createBranchWeights(FallThroughWeight, ExitWeight); 562 LatchBR->setMetadata(LLVMContext::MD_prof, WeightNode); 563 } 564 565 /// Clones the body of the loop L, putting it between \p InsertTop and \p 566 /// InsertBot. 567 /// \param IterNumber The serial number of the iteration currently being 568 /// peeled off. 569 /// \param ExitEdges The exit edges of the original loop. 570 /// \param[out] NewBlocks A list of the blocks in the newly created clone 571 /// \param[out] VMap The value map between the loop and the new clone. 572 /// \param LoopBlocks A helper for DFS-traversal of the loop. 573 /// \param LVMap A value-map that maps instructions from the original loop to 574 /// instructions in the last peeled-off iteration. 575 static void cloneLoopBlocks( 576 Loop *L, unsigned IterNumber, BasicBlock *InsertTop, BasicBlock *InsertBot, 577 SmallVectorImpl<std::pair<BasicBlock *, BasicBlock *>> &ExitEdges, 578 SmallVectorImpl<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks, 579 ValueToValueMapTy &VMap, ValueToValueMapTy &LVMap, DominatorTree *DT, 580 LoopInfo *LI, ArrayRef<MDNode *> LoopLocalNoAliasDeclScopes, 581 ScalarEvolution &SE) { 582 BasicBlock *Header = L->getHeader(); 583 BasicBlock *Latch = L->getLoopLatch(); 584 BasicBlock *PreHeader = L->getLoopPreheader(); 585 586 Function *F = Header->getParent(); 587 LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO(); 588 LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO(); 589 Loop *ParentLoop = L->getParentLoop(); 590 591 // For each block in the original loop, create a new copy, 592 // and update the value map with the newly created values. 593 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { 594 BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, ".peel", F); 595 NewBlocks.push_back(NewBB); 596 597 // If an original block is an immediate child of the loop L, its copy 598 // is a child of a ParentLoop after peeling. If a block is a child of 599 // a nested loop, it is handled in the cloneLoop() call below. 600 if (ParentLoop && LI->getLoopFor(*BB) == L) 601 ParentLoop->addBasicBlockToLoop(NewBB, *LI); 602 603 VMap[*BB] = NewBB; 604 605 // If dominator tree is available, insert nodes to represent cloned blocks. 606 if (DT) { 607 if (Header == *BB) 608 DT->addNewBlock(NewBB, InsertTop); 609 else { 610 DomTreeNode *IDom = DT->getNode(*BB)->getIDom(); 611 // VMap must contain entry for IDom, as the iteration order is RPO. 612 DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDom->getBlock()])); 613 } 614 } 615 } 616 617 { 618 // Identify what other metadata depends on the cloned version. After 619 // cloning, replace the metadata with the corrected version for both 620 // memory instructions and noalias intrinsics. 621 std::string Ext = (Twine("Peel") + Twine(IterNumber)).str(); 622 cloneAndAdaptNoAliasScopes(LoopLocalNoAliasDeclScopes, NewBlocks, 623 Header->getContext(), Ext); 624 } 625 626 // Recursively create the new Loop objects for nested loops, if any, 627 // to preserve LoopInfo. 628 for (Loop *ChildLoop : *L) { 629 cloneLoop(ChildLoop, ParentLoop, VMap, LI, nullptr); 630 } 631 632 // Hook-up the control flow for the newly inserted blocks. 633 // The new header is hooked up directly to the "top", which is either 634 // the original loop preheader (for the first iteration) or the previous 635 // iteration's exiting block (for every other iteration) 636 InsertTop->getTerminator()->setSuccessor(0, cast<BasicBlock>(VMap[Header])); 637 638 // Similarly, for the latch: 639 // The original exiting edge is still hooked up to the loop exit. 640 // The backedge now goes to the "bottom", which is either the loop's real 641 // header (for the last peeled iteration) or the copied header of the next 642 // iteration (for every other iteration) 643 BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]); 644 BranchInst *LatchBR = cast<BranchInst>(NewLatch->getTerminator()); 645 for (unsigned idx = 0, e = LatchBR->getNumSuccessors(); idx < e; ++idx) 646 if (LatchBR->getSuccessor(idx) == Header) { 647 LatchBR->setSuccessor(idx, InsertBot); 648 break; 649 } 650 if (DT) 651 DT->changeImmediateDominator(InsertBot, NewLatch); 652 653 // The new copy of the loop body starts with a bunch of PHI nodes 654 // that pick an incoming value from either the preheader, or the previous 655 // loop iteration. Since this copy is no longer part of the loop, we 656 // resolve this statically: 657 // For the first iteration, we use the value from the preheader directly. 658 // For any other iteration, we replace the phi with the value generated by 659 // the immediately preceding clone of the loop body (which represents 660 // the previous iteration). 661 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 662 PHINode *NewPHI = cast<PHINode>(VMap[&*I]); 663 if (IterNumber == 0) { 664 VMap[&*I] = NewPHI->getIncomingValueForBlock(PreHeader); 665 } else { 666 Value *LatchVal = NewPHI->getIncomingValueForBlock(Latch); 667 Instruction *LatchInst = dyn_cast<Instruction>(LatchVal); 668 if (LatchInst && L->contains(LatchInst)) 669 VMap[&*I] = LVMap[LatchInst]; 670 else 671 VMap[&*I] = LatchVal; 672 } 673 cast<BasicBlock>(VMap[Header])->getInstList().erase(NewPHI); 674 } 675 676 // Fix up the outgoing values - we need to add a value for the iteration 677 // we've just created. Note that this must happen *after* the incoming 678 // values are adjusted, since the value going out of the latch may also be 679 // a value coming into the header. 680 for (auto Edge : ExitEdges) 681 for (PHINode &PHI : Edge.second->phis()) { 682 Value *LatchVal = PHI.getIncomingValueForBlock(Edge.first); 683 Instruction *LatchInst = dyn_cast<Instruction>(LatchVal); 684 if (LatchInst && L->contains(LatchInst)) 685 LatchVal = VMap[LatchVal]; 686 PHI.addIncoming(LatchVal, cast<BasicBlock>(VMap[Edge.first])); 687 SE.forgetValue(&PHI); 688 } 689 690 // LastValueMap is updated with the values for the current loop 691 // which are used the next time this function is called. 692 for (auto KV : VMap) 693 LVMap[KV.first] = KV.second; 694 } 695 696 TargetTransformInfo::PeelingPreferences llvm::gatherPeelingPreferences( 697 Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, 698 Optional<bool> UserAllowPeeling, 699 Optional<bool> UserAllowProfileBasedPeeling, bool UnrollingSpecficValues) { 700 TargetTransformInfo::PeelingPreferences PP; 701 702 // Set the default values. 703 PP.PeelCount = 0; 704 PP.AllowPeeling = true; 705 PP.AllowLoopNestsPeeling = false; 706 PP.PeelProfiledIterations = true; 707 708 // Get the target specifc values. 709 TTI.getPeelingPreferences(L, SE, PP); 710 711 // User specified values using cl::opt. 712 if (UnrollingSpecficValues) { 713 if (UnrollPeelCount.getNumOccurrences() > 0) 714 PP.PeelCount = UnrollPeelCount; 715 if (UnrollAllowPeeling.getNumOccurrences() > 0) 716 PP.AllowPeeling = UnrollAllowPeeling; 717 if (UnrollAllowLoopNestsPeeling.getNumOccurrences() > 0) 718 PP.AllowLoopNestsPeeling = UnrollAllowLoopNestsPeeling; 719 } 720 721 // User specifed values provided by argument. 722 if (UserAllowPeeling) 723 PP.AllowPeeling = *UserAllowPeeling; 724 if (UserAllowProfileBasedPeeling) 725 PP.PeelProfiledIterations = *UserAllowProfileBasedPeeling; 726 727 return PP; 728 } 729 730 /// Peel off the first \p PeelCount iterations of loop \p L. 731 /// 732 /// Note that this does not peel them off as a single straight-line block. 733 /// Rather, each iteration is peeled off separately, and needs to check the 734 /// exit condition. 735 /// For loops that dynamically execute \p PeelCount iterations or less 736 /// this provides a benefit, since the peeled off iterations, which account 737 /// for the bulk of dynamic execution, can be further simplified by scalar 738 /// optimizations. 739 bool llvm::peelLoop(Loop *L, unsigned PeelCount, LoopInfo *LI, 740 ScalarEvolution *SE, DominatorTree &DT, AssumptionCache *AC, 741 bool PreserveLCSSA) { 742 assert(PeelCount > 0 && "Attempt to peel out zero iterations?"); 743 assert(canPeel(L) && "Attempt to peel a loop which is not peelable?"); 744 745 LoopBlocksDFS LoopBlocks(L); 746 LoopBlocks.perform(LI); 747 748 BasicBlock *Header = L->getHeader(); 749 BasicBlock *PreHeader = L->getLoopPreheader(); 750 BasicBlock *Latch = L->getLoopLatch(); 751 SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitEdges; 752 L->getExitEdges(ExitEdges); 753 754 // Remember dominators of blocks we might reach through exits to change them 755 // later. Immediate dominator of such block might change, because we add more 756 // routes which can lead to the exit: we can reach it from the peeled 757 // iterations too. 758 DenseMap<BasicBlock *, BasicBlock *> NonLoopBlocksIDom; 759 for (auto *BB : L->blocks()) { 760 auto *BBDomNode = DT.getNode(BB); 761 SmallVector<BasicBlock *, 16> ChildrenToUpdate; 762 for (auto *ChildDomNode : BBDomNode->children()) { 763 auto *ChildBB = ChildDomNode->getBlock(); 764 if (!L->contains(ChildBB)) 765 ChildrenToUpdate.push_back(ChildBB); 766 } 767 // The new idom of the block will be the nearest common dominator 768 // of all copies of the previous idom. This is equivalent to the 769 // nearest common dominator of the previous idom and the first latch, 770 // which dominates all copies of the previous idom. 771 BasicBlock *NewIDom = DT.findNearestCommonDominator(BB, Latch); 772 for (auto *ChildBB : ChildrenToUpdate) 773 NonLoopBlocksIDom[ChildBB] = NewIDom; 774 } 775 776 Function *F = Header->getParent(); 777 778 // Set up all the necessary basic blocks. It is convenient to split the 779 // preheader into 3 parts - two blocks to anchor the peeled copy of the loop 780 // body, and a new preheader for the "real" loop. 781 782 // Peeling the first iteration transforms. 783 // 784 // PreHeader: 785 // ... 786 // Header: 787 // LoopBody 788 // If (cond) goto Header 789 // Exit: 790 // 791 // into 792 // 793 // InsertTop: 794 // LoopBody 795 // If (!cond) goto Exit 796 // InsertBot: 797 // NewPreHeader: 798 // ... 799 // Header: 800 // LoopBody 801 // If (cond) goto Header 802 // Exit: 803 // 804 // Each following iteration will split the current bottom anchor in two, 805 // and put the new copy of the loop body between these two blocks. That is, 806 // after peeling another iteration from the example above, we'll split 807 // InsertBot, and get: 808 // 809 // InsertTop: 810 // LoopBody 811 // If (!cond) goto Exit 812 // InsertBot: 813 // LoopBody 814 // If (!cond) goto Exit 815 // InsertBot.next: 816 // NewPreHeader: 817 // ... 818 // Header: 819 // LoopBody 820 // If (cond) goto Header 821 // Exit: 822 823 BasicBlock *InsertTop = SplitEdge(PreHeader, Header, &DT, LI); 824 BasicBlock *InsertBot = 825 SplitBlock(InsertTop, InsertTop->getTerminator(), &DT, LI); 826 BasicBlock *NewPreHeader = 827 SplitBlock(InsertBot, InsertBot->getTerminator(), &DT, LI); 828 829 InsertTop->setName(Header->getName() + ".peel.begin"); 830 InsertBot->setName(Header->getName() + ".peel.next"); 831 NewPreHeader->setName(PreHeader->getName() + ".peel.newph"); 832 833 ValueToValueMapTy LVMap; 834 835 // If we have branch weight information, we'll want to update it for the 836 // newly created branches. 837 BranchInst *LatchBR = 838 cast<BranchInst>(cast<BasicBlock>(Latch)->getTerminator()); 839 uint64_t ExitWeight = 0, FallThroughWeight = 0; 840 initBranchWeights(Header, LatchBR, ExitWeight, FallThroughWeight); 841 842 // Identify what noalias metadata is inside the loop: if it is inside the 843 // loop, the associated metadata must be cloned for each iteration. 844 SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes; 845 identifyNoAliasScopesToClone(L->getBlocks(), LoopLocalNoAliasDeclScopes); 846 847 // For each peeled-off iteration, make a copy of the loop. 848 for (unsigned Iter = 0; Iter < PeelCount; ++Iter) { 849 SmallVector<BasicBlock *, 8> NewBlocks; 850 ValueToValueMapTy VMap; 851 852 cloneLoopBlocks(L, Iter, InsertTop, InsertBot, ExitEdges, NewBlocks, 853 LoopBlocks, VMap, LVMap, &DT, LI, 854 LoopLocalNoAliasDeclScopes, *SE); 855 856 // Remap to use values from the current iteration instead of the 857 // previous one. 858 remapInstructionsInBlocks(NewBlocks, VMap); 859 860 // Update IDoms of the blocks reachable through exits. 861 if (Iter == 0) 862 for (auto BBIDom : NonLoopBlocksIDom) 863 DT.changeImmediateDominator(BBIDom.first, 864 cast<BasicBlock>(LVMap[BBIDom.second])); 865 #ifdef EXPENSIVE_CHECKS 866 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 867 #endif 868 869 auto *LatchBRCopy = cast<BranchInst>(VMap[LatchBR]); 870 updateBranchWeights(InsertBot, LatchBRCopy, ExitWeight, FallThroughWeight); 871 // Remove Loop metadata from the latch branch instruction 872 // because it is not the Loop's latch branch anymore. 873 LatchBRCopy->setMetadata(LLVMContext::MD_loop, nullptr); 874 875 InsertTop = InsertBot; 876 InsertBot = SplitBlock(InsertBot, InsertBot->getTerminator(), &DT, LI); 877 InsertBot->setName(Header->getName() + ".peel.next"); 878 879 F->getBasicBlockList().splice(InsertTop->getIterator(), 880 F->getBasicBlockList(), 881 NewBlocks[0]->getIterator(), F->end()); 882 } 883 884 // Now adjust the phi nodes in the loop header to get their initial values 885 // from the last peeled-off iteration instead of the preheader. 886 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 887 PHINode *PHI = cast<PHINode>(I); 888 Value *NewVal = PHI->getIncomingValueForBlock(Latch); 889 Instruction *LatchInst = dyn_cast<Instruction>(NewVal); 890 if (LatchInst && L->contains(LatchInst)) 891 NewVal = LVMap[LatchInst]; 892 893 PHI->setIncomingValueForBlock(NewPreHeader, NewVal); 894 } 895 896 fixupBranchWeights(Header, LatchBR, ExitWeight, FallThroughWeight); 897 898 // Update Metadata for count of peeled off iterations. 899 unsigned AlreadyPeeled = 0; 900 if (auto Peeled = getOptionalIntLoopAttribute(L, PeeledCountMetaData)) 901 AlreadyPeeled = *Peeled; 902 addStringMetadataToLoop(L, PeeledCountMetaData, AlreadyPeeled + PeelCount); 903 904 if (Loop *ParentLoop = L->getParentLoop()) 905 L = ParentLoop; 906 907 // We modified the loop, update SE. 908 SE->forgetTopmostLoop(L); 909 910 #ifdef EXPENSIVE_CHECKS 911 // Finally DomtTree must be correct. 912 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 913 #endif 914 915 // FIXME: Incrementally update loop-simplify 916 simplifyLoop(L, &DT, LI, SE, AC, nullptr, PreserveLCSSA); 917 918 NumPeeled++; 919 920 return true; 921 } 922