lib/Analysis/LoopInfo.cpp

   1 //===- LoopInfo.cpp - Natural Loop Calculator -----------------------------===//
   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 file defines the LoopInfo class that is used to identify natural loops
  10 // and determine the loop depth of various nodes of the CFG.  Note that the
  11 // loops identified may actually be several natural loops that share the same
  12 // header node... not just a single natural loop.
  13 //
  14 //===----------------------------------------------------------------------===//
  15
  16 #include "llvm/Analysis/LoopInfo.h"
  17 #include "llvm/ADT/DepthFirstIterator.h"
  18 #include "llvm/ADT/ScopeExit.h"
  19 #include "llvm/ADT/SmallPtrSet.h"
  20 #include "llvm/Analysis/IVDescriptors.h"
  21 #include "llvm/Analysis/LoopInfoImpl.h"
  22 #include "llvm/Analysis/LoopIterator.h"
  23 #include "llvm/Analysis/MemorySSA.h"
  24 #include "llvm/Analysis/MemorySSAUpdater.h"
  25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
  26 #include "llvm/Analysis/ValueTracking.h"
  27 #include "llvm/Config/llvm-config.h"
  28 #include "llvm/IR/CFG.h"
  29 #include "llvm/IR/Constants.h"
  30 #include "llvm/IR/DebugLoc.h"
  31 #include "llvm/IR/Dominators.h"
  32 #include "llvm/IR/IRPrintingPasses.h"
  33 #include "llvm/IR/Instructions.h"
  34 #include "llvm/IR/LLVMContext.h"
  35 #include "llvm/IR/Metadata.h"
  36 #include "llvm/IR/PassManager.h"
  37 #include "llvm/Support/CommandLine.h"
  38 #include "llvm/Support/Debug.h"
  39 #include "llvm/Support/raw_ostream.h"
  40 #include <algorithm>
  41 using namespace llvm;
  42
  43 // Explicitly instantiate methods in LoopInfoImpl.h for IR-level Loops.
  44 template class llvm::LoopBase<BasicBlock, Loop>;
  45 template class llvm::LoopInfoBase<BasicBlock, Loop>;
  46
  47 // Always verify loopinfo if expensive checking is enabled.
  48 #ifdef EXPENSIVE_CHECKS
  49 bool llvm::VerifyLoopInfo = true;
  50 #else
  51 bool llvm::VerifyLoopInfo = false;
  52 #endif
  53 static cl::opt<bool, true>
  54     VerifyLoopInfoX("verify-loop-info", cl::location(VerifyLoopInfo),
  55                     cl::Hidden, cl::desc("Verify loop info (time consuming)"));
  56
  57 //===----------------------------------------------------------------------===//
  58 // Loop implementation
  59 //
  60
  61 bool Loop::isLoopInvariant(const Value *V) const {
  62   if (const Instruction *I = dyn_cast<Instruction>(V))
  63     return !contains(I);
  64   return true; // All non-instructions are loop invariant
  65 }
  66
  67 bool Loop::hasLoopInvariantOperands(const Instruction *I) const {
  68   return all_of(I->operands(), [this](Value *V) { return isLoopInvariant(V); });
  69 }
  70
  71 bool Loop::makeLoopInvariant(Value *V, bool &Changed, Instruction *InsertPt,
  72                              MemorySSAUpdater *MSSAU) const {
  73   if (Instruction *I = dyn_cast<Instruction>(V))
  74     return makeLoopInvariant(I, Changed, InsertPt, MSSAU);
  75   return true; // All non-instructions are loop-invariant.
  76 }
  77
  78 bool Loop::makeLoopInvariant(Instruction *I, bool &Changed,
  79                              Instruction *InsertPt,
  80                              MemorySSAUpdater *MSSAU) const {
  81   // Test if the value is already loop-invariant.
  82   if (isLoopInvariant(I))
  83     return true;
  84   if (!isSafeToSpeculativelyExecute(I))
  85     return false;
  86   if (I->mayReadFromMemory())
  87     return false;
  88   // EH block instructions are immobile.
  89   if (I->isEHPad())
  90     return false;
  91   // Determine the insertion point, unless one was given.
  92   if (!InsertPt) {
  93     BasicBlock *Preheader = getLoopPreheader();
  94     // Without a preheader, hoisting is not feasible.
  95     if (!Preheader)
  96       return false;
  97     InsertPt = Preheader->getTerminator();
  98   }
  99   // Don't hoist instructions with loop-variant operands.
 100   for (Value *Operand : I->operands())
 101     if (!makeLoopInvariant(Operand, Changed, InsertPt, MSSAU))
 102       return false;
 103
 104   // Hoist.
 105   I->moveBefore(InsertPt);
 106   if (MSSAU)
 107     if (auto *MUD = MSSAU->getMemorySSA()->getMemoryAccess(I))
 108       MSSAU->moveToPlace(MUD, InsertPt->getParent(), MemorySSA::End);
 109
 110   // There is possibility of hoisting this instruction above some arbitrary
 111   // condition. Any metadata defined on it can be control dependent on this
 112   // condition. Conservatively strip it here so that we don't give any wrong
 113   // information to the optimizer.
 114   I->dropUnknownNonDebugMetadata();
 115
 116   Changed = true;
 117   return true;
 118 }
 119
 120 bool Loop::getIncomingAndBackEdge(BasicBlock *&Incoming,
 121                                   BasicBlock *&Backedge) const {
 122   BasicBlock *H = getHeader();
 123
 124   Incoming = nullptr;
 125   Backedge = nullptr;
 126   pred_iterator PI = pred_begin(H);
 127   assert(PI != pred_end(H) && "Loop must have at least one backedge!");
 128   Backedge = *PI++;
 129   if (PI == pred_end(H))
 130     return false; // dead loop
 131   Incoming = *PI++;
 132   if (PI != pred_end(H))
 133     return false; // multiple backedges?
 134
 135   if (contains(Incoming)) {
 136     if (contains(Backedge))
 137       return false;
 138     std::swap(Incoming, Backedge);
 139   } else if (!contains(Backedge))
 140     return false;
 141
 142   assert(Incoming && Backedge && "expected non-null incoming and backedges");
 143   return true;
 144 }
 145
 146 PHINode *Loop::getCanonicalInductionVariable() const {
 147   BasicBlock *H = getHeader();
 148
 149   BasicBlock *Incoming = nullptr, *Backedge = nullptr;
 150   if (!getIncomingAndBackEdge(Incoming, Backedge))
 151     return nullptr;
 152
 153   // Loop over all of the PHI nodes, looking for a canonical indvar.
 154   for (BasicBlock::iterator I = H->begin(); isa<PHINode>(I); ++I) {
 155     PHINode *PN = cast<PHINode>(I);
 156     if (ConstantInt *CI =
 157             dyn_cast<ConstantInt>(PN->getIncomingValueForBlock(Incoming)))
 158       if (CI->isZero())
 159         if (Instruction *Inc =
 160                 dyn_cast<Instruction>(PN->getIncomingValueForBlock(Backedge)))
 161           if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(0) == PN)
 162             if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1)))
 163               if (CI->isOne())
 164                 return PN;
 165   }
 166   return nullptr;
 167 }
 168
 169 /// Get the latch condition instruction.
 170 static ICmpInst *getLatchCmpInst(const Loop &L) {
 171   if (BasicBlock *Latch = L.getLoopLatch())
 172     if (BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator()))
 173       if (BI->isConditional())
 174         return dyn_cast<ICmpInst>(BI->getCondition());
 175
 176   return nullptr;
 177 }
 178
 179 /// Return the final value of the loop induction variable if found.
 180 static Value *findFinalIVValue(const Loop &L, const PHINode &IndVar,
 181                                const Instruction &StepInst) {
 182   ICmpInst *LatchCmpInst = getLatchCmpInst(L);
 183   if (!LatchCmpInst)
 184     return nullptr;
 185
 186   Value *Op0 = LatchCmpInst->getOperand(0);
 187   Value *Op1 = LatchCmpInst->getOperand(1);
 188   if (Op0 == &IndVar || Op0 == &StepInst)
 189     return Op1;
 190
 191   if (Op1 == &IndVar || Op1 == &StepInst)
 192     return Op0;
 193
 194   return nullptr;
 195 }
 196
 197 Optional<Loop::LoopBounds> Loop::LoopBounds::getBounds(const Loop &L,
 198                                                        PHINode &IndVar,
 199                                                        ScalarEvolution &SE) {
 200   InductionDescriptor IndDesc;
 201   if (!InductionDescriptor::isInductionPHI(&IndVar, &L, &SE, IndDesc))
 202     return None;
 203
 204   Value *InitialIVValue = IndDesc.getStartValue();
 205   Instruction *StepInst = IndDesc.getInductionBinOp();
 206   if (!InitialIVValue || !StepInst)
 207     return None;
 208
 209   const SCEV *Step = IndDesc.getStep();
 210   Value *StepInstOp1 = StepInst->getOperand(1);
 211   Value *StepInstOp0 = StepInst->getOperand(0);
 212   Value *StepValue = nullptr;
 213   if (SE.getSCEV(StepInstOp1) == Step)
 214     StepValue = StepInstOp1;
 215   else if (SE.getSCEV(StepInstOp0) == Step)
 216     StepValue = StepInstOp0;
 217
 218   Value *FinalIVValue = findFinalIVValue(L, IndVar, *StepInst);
 219   if (!FinalIVValue)
 220     return None;
 221
 222   return LoopBounds(L, *InitialIVValue, *StepInst, StepValue, *FinalIVValue,
 223                     SE);
 224 }
 225
 226 using Direction = Loop::LoopBounds::Direction;
 227
 228 ICmpInst::Predicate Loop::LoopBounds::getCanonicalPredicate() const {
 229   BasicBlock *Latch = L.getLoopLatch();
 230   assert(Latch && "Expecting valid latch");
 231
 232   BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator());
 233   assert(BI && BI->isConditional() && "Expecting conditional latch branch");
 234
 235   ICmpInst *LatchCmpInst = dyn_cast<ICmpInst>(BI->getCondition());
 236   assert(LatchCmpInst &&
 237          "Expecting the latch compare instruction to be a CmpInst");
 238
 239   // Need to inverse the predicate when first successor is not the loop
 240   // header
 241   ICmpInst::Predicate Pred = (BI->getSuccessor(0) == L.getHeader())
 242                                  ? LatchCmpInst->getPredicate()
 243                                  : LatchCmpInst->getInversePredicate();
 244
 245   if (LatchCmpInst->getOperand(0) == &getFinalIVValue())
 246     Pred = ICmpInst::getSwappedPredicate(Pred);
 247
 248   // Need to flip strictness of the predicate when the latch compare instruction
 249   // is not using StepInst
 250   if (LatchCmpInst->getOperand(0) == &getStepInst() ||
 251       LatchCmpInst->getOperand(1) == &getStepInst())
 252     return Pred;
 253
 254   // Cannot flip strictness of NE and EQ
 255   if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
 256     return ICmpInst::getFlippedStrictnessPredicate(Pred);
 257
 258   Direction D = getDirection();
 259   if (D == Direction::Increasing)
 260     return ICmpInst::ICMP_SLT;
 261
 262   if (D == Direction::Decreasing)
 263     return ICmpInst::ICMP_SGT;
 264
 265   // If cannot determine the direction, then unable to find the canonical
 266   // predicate
 267   return ICmpInst::BAD_ICMP_PREDICATE;
 268 }
 269
 270 Direction Loop::LoopBounds::getDirection() const {
 271   if (const SCEVAddRecExpr *StepAddRecExpr =
 272           dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&getStepInst())))
 273     if (const SCEV *StepRecur = StepAddRecExpr->getStepRecurrence(SE)) {
 274       if (SE.isKnownPositive(StepRecur))
 275         return Direction::Increasing;
 276       if (SE.isKnownNegative(StepRecur))
 277         return Direction::Decreasing;
 278     }
 279
 280   return Direction::Unknown;
 281 }
 282
 283 Optional<Loop::LoopBounds> Loop::getBounds(ScalarEvolution &SE) const {
 284   if (PHINode *IndVar = getInductionVariable(SE))
 285     return LoopBounds::getBounds(*this, *IndVar, SE);
 286
 287   return None;
 288 }
 289
 290 PHINode *Loop::getInductionVariable(ScalarEvolution &SE) const {
 291   if (!isLoopSimplifyForm())
 292     return nullptr;
 293
 294   BasicBlock *Header = getHeader();
 295   assert(Header && "Expected a valid loop header");
 296   ICmpInst *CmpInst = getLatchCmpInst(*this);
 297   if (!CmpInst)
 298     return nullptr;
 299
 300   Instruction *LatchCmpOp0 = dyn_cast<Instruction>(CmpInst->getOperand(0));
 301   Instruction *LatchCmpOp1 = dyn_cast<Instruction>(CmpInst->getOperand(1));
 302
 303   for (PHINode &IndVar : Header->phis()) {
 304     InductionDescriptor IndDesc;
 305     if (!InductionDescriptor::isInductionPHI(&IndVar, this, &SE, IndDesc))
 306       continue;
 307
 308     Instruction *StepInst = IndDesc.getInductionBinOp();
 309
 310     // case 1:
 311     // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
 312     // StepInst = IndVar + step
 313     // cmp = StepInst < FinalValue
 314     if (StepInst == LatchCmpOp0 || StepInst == LatchCmpOp1)
 315       return &IndVar;
 316
 317     // case 2:
 318     // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
 319     // StepInst = IndVar + step
 320     // cmp = IndVar < FinalValue
 321     if (&IndVar == LatchCmpOp0 || &IndVar == LatchCmpOp1)
 322       return &IndVar;
 323   }
 324
 325   return nullptr;
 326 }
 327
 328 bool Loop::getInductionDescriptor(ScalarEvolution &SE,
 329                                   InductionDescriptor &IndDesc) const {
 330   if (PHINode *IndVar = getInductionVariable(SE))
 331     return InductionDescriptor::isInductionPHI(IndVar, this, &SE, IndDesc);
 332
 333   return false;
 334 }
 335
 336 bool Loop::isAuxiliaryInductionVariable(PHINode &AuxIndVar,
 337                                         ScalarEvolution &SE) const {
 338   // Located in the loop header
 339   BasicBlock *Header = getHeader();
 340   if (AuxIndVar.getParent() != Header)
 341     return false;
 342
 343   // No uses outside of the loop
 344   for (User *U : AuxIndVar.users())
 345     if (const Instruction *I = dyn_cast<Instruction>(U))
 346       if (!contains(I))
 347         return false;
 348
 349   InductionDescriptor IndDesc;
 350   if (!InductionDescriptor::isInductionPHI(&AuxIndVar, this, &SE, IndDesc))
 351     return false;
 352
 353   // The step instruction opcode should be add or sub.
 354   if (IndDesc.getInductionOpcode() != Instruction::Add &&
 355       IndDesc.getInductionOpcode() != Instruction::Sub)
 356     return false;
 357
 358   // Incremented by a loop invariant step for each loop iteration
 359   return SE.isLoopInvariant(IndDesc.getStep(), this);
 360 }
 361
 362 BranchInst *Loop::getLoopGuardBranch() const {
 363   if (!isLoopSimplifyForm())
 364     return nullptr;
 365
 366   BasicBlock *Preheader = getLoopPreheader();
 367   BasicBlock *Latch = getLoopLatch();
 368   assert(Preheader && Latch &&
 369          "Expecting a loop with valid preheader and latch");
 370
 371   // Loop should be in rotate form.
 372   if (!isLoopExiting(Latch))
 373     return nullptr;
 374
 375   // Disallow loops with more than one unique exit block, as we do not verify
 376   // that GuardOtherSucc post dominates all exit blocks.
 377   BasicBlock *ExitFromLatch = getUniqueExitBlock();
 378   if (!ExitFromLatch)
 379     return nullptr;
 380
 381   BasicBlock *ExitFromLatchSucc = ExitFromLatch->getUniqueSuccessor();
 382   if (!ExitFromLatchSucc)
 383     return nullptr;
 384
 385   BasicBlock *GuardBB = Preheader->getUniquePredecessor();
 386   if (!GuardBB)
 387     return nullptr;
 388
 389   assert(GuardBB->getTerminator() && "Expecting valid guard terminator");
 390
 391   BranchInst *GuardBI = dyn_cast<BranchInst>(GuardBB->getTerminator());
 392   if (!GuardBI || GuardBI->isUnconditional())
 393     return nullptr;
 394
 395   BasicBlock *GuardOtherSucc = (GuardBI->getSuccessor(0) == Preheader)
 396                                    ? GuardBI->getSuccessor(1)
 397                                    : GuardBI->getSuccessor(0);
 398   return (GuardOtherSucc == ExitFromLatchSucc) ? GuardBI : nullptr;
 399 }
 400
 401 bool Loop::isCanonical(ScalarEvolution &SE) const {
 402   InductionDescriptor IndDesc;
 403   if (!getInductionDescriptor(SE, IndDesc))
 404     return false;
 405
 406   ConstantInt *Init = dyn_cast_or_null<ConstantInt>(IndDesc.getStartValue());
 407   if (!Init || !Init->isZero())
 408     return false;
 409
 410   if (IndDesc.getInductionOpcode() != Instruction::Add)
 411     return false;
 412
 413   ConstantInt *Step = IndDesc.getConstIntStepValue();
 414   if (!Step || !Step->isOne())
 415     return false;
 416
 417   return true;
 418 }
 419
 420 // Check that 'BB' doesn't have any uses outside of the 'L'
 421 static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB,
 422                                DominatorTree &DT) {
 423   for (const Instruction &I : BB) {
 424     // Tokens can't be used in PHI nodes and live-out tokens prevent loop
 425     // optimizations, so for the purposes of considered LCSSA form, we
 426     // can ignore them.
 427     if (I.getType()->isTokenTy())
 428       continue;
 429
 430     for (const Use &U : I.uses()) {
 431       const Instruction *UI = cast<Instruction>(U.getUser());
 432       const BasicBlock *UserBB = UI->getParent();
 433       if (const PHINode *P = dyn_cast<PHINode>(UI))
 434         UserBB = P->getIncomingBlock(U);
 435
 436       // Check the current block, as a fast-path, before checking whether
 437       // the use is anywhere in the loop.  Most values are used in the same
 438       // block they are defined in.  Also, blocks not reachable from the
 439       // entry are special; uses in them don't need to go through PHIs.
 440       if (UserBB != &BB && !L.contains(UserBB) &&
 441           DT.isReachableFromEntry(UserBB))
 442         return false;
 443     }
 444   }
 445   return true;
 446 }
 447
 448 bool Loop::isLCSSAForm(DominatorTree &DT) const {
 449   // For each block we check that it doesn't have any uses outside of this loop.
 450   return all_of(this->blocks(), [&](const BasicBlock *BB) {
 451     return isBlockInLCSSAForm(*this, *BB, DT);
 452   });
 453 }
 454
 455 bool Loop::isRecursivelyLCSSAForm(DominatorTree &DT, const LoopInfo &LI) const {
 456   // For each block we check that it doesn't have any uses outside of its
 457   // innermost loop. This process will transitively guarantee that the current
 458   // loop and all of the nested loops are in LCSSA form.
 459   return all_of(this->blocks(), [&](const BasicBlock *BB) {
 460     return isBlockInLCSSAForm(*LI.getLoopFor(BB), *BB, DT);
 461   });
 462 }
 463
 464 bool Loop::isLoopSimplifyForm() const {
 465   // Normal-form loops have a preheader, a single backedge, and all of their
 466   // exits have all their predecessors inside the loop.
 467   return getLoopPreheader() && getLoopLatch() && hasDedicatedExits();
 468 }
 469
 470 // Routines that reform the loop CFG and split edges often fail on indirectbr.
 471 bool Loop::isSafeToClone() const {
 472   // Return false if any loop blocks contain indirectbrs, or there are any calls
 473   // to noduplicate functions.
 474   // FIXME: it should be ok to clone CallBrInst's if we correctly update the
 475   // operand list to reflect the newly cloned labels.
 476   for (BasicBlock *BB : this->blocks()) {
 477     if (isa<IndirectBrInst>(BB->getTerminator()) ||
 478         isa<CallBrInst>(BB->getTerminator()))
 479       return false;
 480
 481     for (Instruction &I : *BB)
 482       if (auto CS = CallSite(&I))
 483         if (CS.cannotDuplicate())
 484           return false;
 485   }
 486   return true;
 487 }
 488
 489 MDNode *Loop::getLoopID() const {
 490   MDNode *LoopID = nullptr;
 491
 492   // Go through the latch blocks and check the terminator for the metadata.
 493   SmallVector<BasicBlock *, 4> LatchesBlocks;
 494   getLoopLatches(LatchesBlocks);
 495   for (BasicBlock *BB : LatchesBlocks) {
 496     Instruction *TI = BB->getTerminator();
 497     MDNode *MD = TI->getMetadata(LLVMContext::MD_loop);
 498
 499     if (!MD)
 500       return nullptr;
 501
 502     if (!LoopID)
 503       LoopID = MD;
 504     else if (MD != LoopID)
 505       return nullptr;
 506   }
 507   if (!LoopID || LoopID->getNumOperands() == 0 ||
 508       LoopID->getOperand(0) != LoopID)
 509     return nullptr;
 510   return LoopID;
 511 }
 512
 513 void Loop::setLoopID(MDNode *LoopID) const {
 514   assert((!LoopID || LoopID->getNumOperands() > 0) &&
 515          "Loop ID needs at least one operand");
 516   assert((!LoopID || LoopID->getOperand(0) == LoopID) &&
 517          "Loop ID should refer to itself");
 518
 519   SmallVector<BasicBlock *, 4> LoopLatches;
 520   getLoopLatches(LoopLatches);
 521   for (BasicBlock *BB : LoopLatches)
 522     BB->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID);
 523 }
 524
 525 void Loop::setLoopAlreadyUnrolled() {
 526   LLVMContext &Context = getHeader()->getContext();
 527
 528   MDNode *DisableUnrollMD =
 529       MDNode::get(Context, MDString::get(Context, "llvm.loop.unroll.disable"));
 530   MDNode *LoopID = getLoopID();
 531   MDNode *NewLoopID = makePostTransformationMetadata(
 532       Context, LoopID, {"llvm.loop.unroll."}, {DisableUnrollMD});
 533   setLoopID(NewLoopID);
 534 }
 535
 536 bool Loop::isAnnotatedParallel() const {
 537   MDNode *DesiredLoopIdMetadata = getLoopID();
 538
 539   if (!DesiredLoopIdMetadata)
 540     return false;
 541
 542   MDNode *ParallelAccesses =
 543       findOptionMDForLoop(this, "llvm.loop.parallel_accesses");
 544   SmallPtrSet<MDNode *, 4>
 545       ParallelAccessGroups; // For scalable 'contains' check.
 546   if (ParallelAccesses) {
 547     for (const MDOperand &MD : drop_begin(ParallelAccesses->operands(), 1)) {
 548       MDNode *AccGroup = cast<MDNode>(MD.get());
 549       assert(isValidAsAccessGroup(AccGroup) &&
 550              "List item must be an access group");
 551       ParallelAccessGroups.insert(AccGroup);
 552     }
 553   }
 554
 555   // The loop branch contains the parallel loop metadata. In order to ensure
 556   // that any parallel-loop-unaware optimization pass hasn't added loop-carried
 557   // dependencies (thus converted the loop back to a sequential loop), check
 558   // that all the memory instructions in the loop belong to an access group that
 559   // is parallel to this loop.
 560   for (BasicBlock *BB : this->blocks()) {
 561     for (Instruction &I : *BB) {
 562       if (!I.mayReadOrWriteMemory())
 563         continue;
 564
 565       if (MDNode *AccessGroup = I.getMetadata(LLVMContext::MD_access_group)) {
 566         auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool {
 567           if (AG->getNumOperands() == 0) {
 568             assert(isValidAsAccessGroup(AG) && "Item must be an access group");
 569             return ParallelAccessGroups.count(AG);
 570           }
 571
 572           for (const MDOperand &AccessListItem : AG->operands()) {
 573             MDNode *AccGroup = cast<MDNode>(AccessListItem.get());
 574             assert(isValidAsAccessGroup(AccGroup) &&
 575                    "List item must be an access group");
 576             if (ParallelAccessGroups.count(AccGroup))
 577               return true;
 578           }
 579           return false;
 580         };
 581
 582         if (ContainsAccessGroup(AccessGroup))
 583           continue;
 584       }
 585
 586       // The memory instruction can refer to the loop identifier metadata
 587       // directly or indirectly through another list metadata (in case of
 588       // nested parallel loops). The loop identifier metadata refers to
 589       // itself so we can check both cases with the same routine.
 590       MDNode *LoopIdMD =
 591           I.getMetadata(LLVMContext::MD_mem_parallel_loop_access);
 592
 593       if (!LoopIdMD)
 594         return false;
 595
 596       bool LoopIdMDFound = false;
 597       for (const MDOperand &MDOp : LoopIdMD->operands()) {
 598         if (MDOp == DesiredLoopIdMetadata) {
 599           LoopIdMDFound = true;
 600           break;
 601         }
 602       }
 603
 604       if (!LoopIdMDFound)
 605         return false;
 606     }
 607   }
 608   return true;
 609 }
 610
 611 DebugLoc Loop::getStartLoc() const { return getLocRange().getStart(); }
 612
 613 Loop::LocRange Loop::getLocRange() const {
 614   // If we have a debug location in the loop ID, then use it.
 615   if (MDNode *LoopID = getLoopID()) {
 616     DebugLoc Start;
 617     // We use the first DebugLoc in the header as the start location of the loop
 618     // and if there is a second DebugLoc in the header we use it as end location
 619     // of the loop.
 620     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
 621       if (DILocation *L = dyn_cast<DILocation>(LoopID->getOperand(i))) {
 622         if (!Start)
 623           Start = DebugLoc(L);
 624         else
 625           return LocRange(Start, DebugLoc(L));
 626       }
 627     }
 628
 629     if (Start)
 630       return LocRange(Start);
 631   }
 632
 633   // Try the pre-header first.
 634   if (BasicBlock *PHeadBB = getLoopPreheader())
 635     if (DebugLoc DL = PHeadBB->getTerminator()->getDebugLoc())
 636       return LocRange(DL);
 637
 638   // If we have no pre-header or there are no instructions with debug
 639   // info in it, try the header.
 640   if (BasicBlock *HeadBB = getHeader())
 641     return LocRange(HeadBB->getTerminator()->getDebugLoc());
 642
 643   return LocRange();
 644 }
 645
 646 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 647 LLVM_DUMP_METHOD void Loop::dump() const { print(dbgs()); }
 648
 649 LLVM_DUMP_METHOD void Loop::dumpVerbose() const {
 650   print(dbgs(), /*Depth=*/0, /*Verbose=*/true);
 651 }
 652 #endif
 653
 654 //===----------------------------------------------------------------------===//
 655 // UnloopUpdater implementation
 656 //
 657
 658 namespace {
 659 /// Find the new parent loop for all blocks within the "unloop" whose last
 660 /// backedges has just been removed.
 661 class UnloopUpdater {
 662   Loop &Unloop;
 663   LoopInfo *LI;
 664
 665   LoopBlocksDFS DFS;
 666
 667   // Map unloop's immediate subloops to their nearest reachable parents. Nested
 668   // loops within these subloops will not change parents. However, an immediate
 669   // subloop's new parent will be the nearest loop reachable from either its own
 670   // exits *or* any of its nested loop's exits.
 671   DenseMap<Loop *, Loop *> SubloopParents;
 672
 673   // Flag the presence of an irreducible backedge whose destination is a block
 674   // directly contained by the original unloop.
 675   bool FoundIB;
 676
 677 public:
 678   UnloopUpdater(Loop *UL, LoopInfo *LInfo)
 679       : Unloop(*UL), LI(LInfo), DFS(UL), FoundIB(false) {}
 680
 681   void updateBlockParents();
 682
 683   void removeBlocksFromAncestors();
 684
 685   void updateSubloopParents();
 686
 687 protected:
 688   Loop *getNearestLoop(BasicBlock *BB, Loop *BBLoop);
 689 };
 690 } // end anonymous namespace
 691
 692 /// Update the parent loop for all blocks that are directly contained within the
 693 /// original "unloop".
 694 void UnloopUpdater::updateBlockParents() {
 695   if (Unloop.getNumBlocks()) {
 696     // Perform a post order CFG traversal of all blocks within this loop,
 697     // propagating the nearest loop from successors to predecessors.
 698     LoopBlocksTraversal Traversal(DFS, LI);
 699     for (BasicBlock *POI : Traversal) {
 700
 701       Loop *L = LI->getLoopFor(POI);
 702       Loop *NL = getNearestLoop(POI, L);
 703
 704       if (NL != L) {
 705         // For reducible loops, NL is now an ancestor of Unloop.
 706         assert((NL != &Unloop && (!NL || NL->contains(&Unloop))) &&
 707                "uninitialized successor");
 708         LI->changeLoopFor(POI, NL);
 709       } else {
 710         // Or the current block is part of a subloop, in which case its parent
 711         // is unchanged.
 712         assert((FoundIB || Unloop.contains(L)) && "uninitialized successor");
 713       }
 714     }
 715   }
 716   // Each irreducible loop within the unloop induces a round of iteration using
 717   // the DFS result cached by Traversal.
 718   bool Changed = FoundIB;
 719   for (unsigned NIters = 0; Changed; ++NIters) {
 720     assert(NIters < Unloop.getNumBlocks() && "runaway iterative algorithm");
 721
 722     // Iterate over the postorder list of blocks, propagating the nearest loop
 723     // from successors to predecessors as before.
 724     Changed = false;
 725     for (LoopBlocksDFS::POIterator POI = DFS.beginPostorder(),
 726                                    POE = DFS.endPostorder();
 727          POI != POE; ++POI) {
 728
 729       Loop *L = LI->getLoopFor(*POI);
 730       Loop *NL = getNearestLoop(*POI, L);
 731       if (NL != L) {
 732         assert(NL != &Unloop && (!NL || NL->contains(&Unloop)) &&
 733                "uninitialized successor");
 734         LI->changeLoopFor(*POI, NL);
 735         Changed = true;
 736       }
 737     }
 738   }
 739 }
 740
 741 /// Remove unloop's blocks from all ancestors below their new parents.
 742 void UnloopUpdater::removeBlocksFromAncestors() {
 743   // Remove all unloop's blocks (including those in nested subloops) from
 744   // ancestors below the new parent loop.
 745   for (Loop::block_iterator BI = Unloop.block_begin(), BE = Unloop.block_end();
 746        BI != BE; ++BI) {
 747     Loop *OuterParent = LI->getLoopFor(*BI);
 748     if (Unloop.contains(OuterParent)) {
 749       while (OuterParent->getParentLoop() != &Unloop)
 750         OuterParent = OuterParent->getParentLoop();
 751       OuterParent = SubloopParents[OuterParent];
 752     }
 753     // Remove blocks from former Ancestors except Unloop itself which will be
 754     // deleted.
 755     for (Loop *OldParent = Unloop.getParentLoop(); OldParent != OuterParent;
 756          OldParent = OldParent->getParentLoop()) {
 757       assert(OldParent && "new loop is not an ancestor of the original");
 758       OldParent->removeBlockFromLoop(*BI);
 759     }
 760   }
 761 }
 762
 763 /// Update the parent loop for all subloops directly nested within unloop.
 764 void UnloopUpdater::updateSubloopParents() {
 765   while (!Unloop.empty()) {
 766     Loop *Subloop = *std::prev(Unloop.end());
 767     Unloop.removeChildLoop(std::prev(Unloop.end()));
 768
 769     assert(SubloopParents.count(Subloop) && "DFS failed to visit subloop");
 770     if (Loop *Parent = SubloopParents[Subloop])
 771       Parent->addChildLoop(Subloop);
 772     else
 773       LI->addTopLevelLoop(Subloop);
 774   }
 775 }
 776
 777 /// Return the nearest parent loop among this block's successors. If a successor
 778 /// is a subloop header, consider its parent to be the nearest parent of the
 779 /// subloop's exits.
 780 ///
 781 /// For subloop blocks, simply update SubloopParents and return NULL.
 782 Loop *UnloopUpdater::getNearestLoop(BasicBlock *BB, Loop *BBLoop) {
 783
 784   // Initially for blocks directly contained by Unloop, NearLoop == Unloop and
 785   // is considered uninitialized.
 786   Loop *NearLoop = BBLoop;
 787
 788   Loop *Subloop = nullptr;
 789   if (NearLoop != &Unloop && Unloop.contains(NearLoop)) {
 790     Subloop = NearLoop;
 791     // Find the subloop ancestor that is directly contained within Unloop.
 792     while (Subloop->getParentLoop() != &Unloop) {
 793       Subloop = Subloop->getParentLoop();
 794       assert(Subloop && "subloop is not an ancestor of the original loop");
 795     }
 796     // Get the current nearest parent of the Subloop exits, initially Unloop.
 797     NearLoop = SubloopParents.insert({Subloop, &Unloop}).first->second;
 798   }
 799
 800   succ_iterator I = succ_begin(BB), E = succ_end(BB);
 801   if (I == E) {
 802     assert(!Subloop && "subloop blocks must have a successor");
 803     NearLoop = nullptr; // unloop blocks may now exit the function.
 804   }
 805   for (; I != E; ++I) {
 806     if (*I == BB)
 807       continue; // self loops are uninteresting
 808
 809     Loop *L = LI->getLoopFor(*I);
 810     if (L == &Unloop) {
 811       // This successor has not been processed. This path must lead to an
 812       // irreducible backedge.
 813       assert((FoundIB || !DFS.hasPostorder(*I)) && "should have seen IB");
 814       FoundIB = true;
 815     }
 816     if (L != &Unloop && Unloop.contains(L)) {
 817       // Successor is in a subloop.
 818       if (Subloop)
 819         continue; // Branching within subloops. Ignore it.
 820
 821       // BB branches from the original into a subloop header.
 822       assert(L->getParentLoop() == &Unloop && "cannot skip into nested loops");
 823
 824       // Get the current nearest parent of the Subloop's exits.
 825       L = SubloopParents[L];
 826       // L could be Unloop if the only exit was an irreducible backedge.
 827     }
 828     if (L == &Unloop) {
 829       continue;
 830     }
 831     // Handle critical edges from Unloop into a sibling loop.
 832     if (L && !L->contains(&Unloop)) {
 833       L = L->getParentLoop();
 834     }
 835     // Remember the nearest parent loop among successors or subloop exits.
 836     if (NearLoop == &Unloop || !NearLoop || NearLoop->contains(L))
 837       NearLoop = L;
 838   }
 839   if (Subloop) {
 840     SubloopParents[Subloop] = NearLoop;
 841     return BBLoop;
 842   }
 843   return NearLoop;
 844 }
 845
 846 LoopInfo::LoopInfo(const DomTreeBase<BasicBlock> &DomTree) { analyze(DomTree); }
 847
 848 bool LoopInfo::invalidate(Function &F, const PreservedAnalyses &PA,
 849                           FunctionAnalysisManager::Invalidator &) {
 850   // Check whether the analysis, all analyses on functions, or the function's
 851   // CFG have been preserved.
 852   auto PAC = PA.getChecker<LoopAnalysis>();
 853   return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() ||
 854            PAC.preservedSet<CFGAnalyses>());
 855 }
 856
 857 void LoopInfo::erase(Loop *Unloop) {
 858   assert(!Unloop->isInvalid() && "Loop has already been erased!");
 859
 860   auto InvalidateOnExit = make_scope_exit([&]() { destroy(Unloop); });
 861
 862   // First handle the special case of no parent loop to simplify the algorithm.
 863   if (!Unloop->getParentLoop()) {
 864     // Since BBLoop had no parent, Unloop blocks are no longer in a loop.
 865     for (Loop::block_iterator I = Unloop->block_begin(),
 866                               E = Unloop->block_end();
 867          I != E; ++I) {
 868
 869       // Don't reparent blocks in subloops.
 870       if (getLoopFor(*I) != Unloop)
 871         continue;
 872
 873       // Blocks no longer have a parent but are still referenced by Unloop until
 874       // the Unloop object is deleted.
 875       changeLoopFor(*I, nullptr);
 876     }
 877
 878     // Remove the loop from the top-level LoopInfo object.
 879     for (iterator I = begin();; ++I) {
 880       assert(I != end() && "Couldn't find loop");
 881       if (*I == Unloop) {
 882         removeLoop(I);
 883         break;
 884       }
 885     }
 886
 887     // Move all of the subloops to the top-level.
 888     while (!Unloop->empty())
 889       addTopLevelLoop(Unloop->removeChildLoop(std::prev(Unloop->end())));
 890
 891     return;
 892   }
 893
 894   // Update the parent loop for all blocks within the loop. Blocks within
 895   // subloops will not change parents.
 896   UnloopUpdater Updater(Unloop, this);
 897   Updater.updateBlockParents();
 898
 899   // Remove blocks from former ancestor loops.
 900   Updater.removeBlocksFromAncestors();
 901
 902   // Add direct subloops as children in their new parent loop.
 903   Updater.updateSubloopParents();
 904
 905   // Remove unloop from its parent loop.
 906   Loop *ParentLoop = Unloop->getParentLoop();
 907   for (Loop::iterator I = ParentLoop->begin();; ++I) {
 908     assert(I != ParentLoop->end() && "Couldn't find loop");
 909     if (*I == Unloop) {
 910       ParentLoop->removeChildLoop(I);
 911       break;
 912     }
 913   }
 914 }
 915
 916 AnalysisKey LoopAnalysis::Key;
 917
 918 LoopInfo LoopAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
 919   // FIXME: Currently we create a LoopInfo from scratch for every function.
 920   // This may prove to be too wasteful due to deallocating and re-allocating
 921   // memory each time for the underlying map and vector datastructures. At some
 922   // point it may prove worthwhile to use a freelist and recycle LoopInfo
 923   // objects. I don't want to add that kind of complexity until the scope of
 924   // the problem is better understood.
 925   LoopInfo LI;
 926   LI.analyze(AM.getResult<DominatorTreeAnalysis>(F));
 927   return LI;
 928 }
 929
 930 PreservedAnalyses LoopPrinterPass::run(Function &F,
 931                                        FunctionAnalysisManager &AM) {
 932   AM.getResult<LoopAnalysis>(F).print(OS);
 933   return PreservedAnalyses::all();
 934 }
 935
 936 void llvm::printLoop(Loop &L, raw_ostream &OS, const std::string &Banner) {
 937
 938   if (forcePrintModuleIR()) {
 939     // handling -print-module-scope
 940     OS << Banner << " (loop: ";
 941     L.getHeader()->printAsOperand(OS, false);
 942     OS << ")\n";
 943
 944     // printing whole module
 945     OS << *L.getHeader()->getModule();
 946     return;
 947   }
 948
 949   OS << Banner;
 950
 951   auto *PreHeader = L.getLoopPreheader();
 952   if (PreHeader) {
 953     OS << "\n; Preheader:";
 954     PreHeader->print(OS);
 955     OS << "\n; Loop:";
 956   }
 957
 958   for (auto *Block : L.blocks())
 959     if (Block)
 960       Block->print(OS);
 961     else
 962       OS << "Printing <null> block";
 963
 964   SmallVector<BasicBlock *, 8> ExitBlocks;
 965   L.getExitBlocks(ExitBlocks);
 966   if (!ExitBlocks.empty()) {
 967     OS << "\n; Exit blocks";
 968     for (auto *Block : ExitBlocks)
 969       if (Block)
 970         Block->print(OS);
 971       else
 972         OS << "Printing <null> block";
 973   }
 974 }
 975
 976 MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) {
 977   // No loop metadata node, no loop properties.
 978   if (!LoopID)
 979     return nullptr;
 980
 981   // First operand should refer to the metadata node itself, for legacy reasons.
 982   assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
 983   assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
 984
 985   // Iterate over the metdata node operands and look for MDString metadata.
 986   for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
 987     MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
 988     if (!MD || MD->getNumOperands() < 1)
 989       continue;
 990     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
 991     if (!S)
 992       continue;
 993     // Return the operand node if MDString holds expected metadata.
 994     if (Name.equals(S->getString()))
 995       return MD;
 996   }
 997
 998   // Loop property not found.
 999   return nullptr;
1000 }
1001
1002 MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) {
1003   return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
1004 }
1005
1006 bool llvm::isValidAsAccessGroup(MDNode *Node) {
1007   return Node->getNumOperands() == 0 && Node->isDistinct();
1008 }
1009
1010 MDNode *llvm::makePostTransformationMetadata(LLVMContext &Context,
1011                                              MDNode *OrigLoopID,
1012                                              ArrayRef<StringRef> RemovePrefixes,
1013                                              ArrayRef<MDNode *> AddAttrs) {
1014   // First remove any existing loop metadata related to this transformation.
1015   SmallVector<Metadata *, 4> MDs;
1016
1017   // Reserve first location for self reference to the LoopID metadata node.
1018   TempMDTuple TempNode = MDNode::getTemporary(Context, None);
1019   MDs.push_back(TempNode.get());
1020
1021   // Remove metadata for the transformation that has been applied or that became
1022   // outdated.
1023   if (OrigLoopID) {
1024     for (unsigned i = 1, ie = OrigLoopID->getNumOperands(); i < ie; ++i) {
1025       bool IsVectorMetadata = false;
1026       Metadata *Op = OrigLoopID->getOperand(i);
1027       if (MDNode *MD = dyn_cast<MDNode>(Op)) {
1028         const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1029         if (S)
1030           IsVectorMetadata =
1031               llvm::any_of(RemovePrefixes, [S](StringRef Prefix) -> bool {
1032                 return S->getString().startswith(Prefix);
1033               });
1034       }
1035       if (!IsVectorMetadata)
1036         MDs.push_back(Op);
1037     }
1038   }
1039
1040   // Add metadata to avoid reapplying a transformation, such as
1041   // llvm.loop.unroll.disable and llvm.loop.isvectorized.
1042   MDs.append(AddAttrs.begin(), AddAttrs.end());
1043
1044   MDNode *NewLoopID = MDNode::getDistinct(Context, MDs);
1045   // Replace the temporary node with a self-reference.
1046   NewLoopID->replaceOperandWith(0, NewLoopID);
1047   return NewLoopID;
1048 }
1049
1050 //===----------------------------------------------------------------------===//
1051 // LoopInfo implementation
1052 //
1053
1054 char LoopInfoWrapperPass::ID = 0;
1055 INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops", "Natural Loop Information",
1056                       true, true)
1057 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1058 INITIALIZE_PASS_END(LoopInfoWrapperPass, "loops", "Natural Loop Information",
1059                     true, true)
1060
1061 bool LoopInfoWrapperPass::runOnFunction(Function &) {
1062   releaseMemory();
1063   LI.analyze(getAnalysis<DominatorTreeWrapperPass>().getDomTree());
1064   return false;
1065 }
1066
1067 void LoopInfoWrapperPass::verifyAnalysis() const {
1068   // LoopInfoWrapperPass is a FunctionPass, but verifying every loop in the
1069   // function each time verifyAnalysis is called is very expensive. The
1070   // -verify-loop-info option can enable this. In order to perform some
1071   // checking by default, LoopPass has been taught to call verifyLoop manually
1072   // during loop pass sequences.
1073   if (VerifyLoopInfo) {
1074     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1075     LI.verify(DT);
1076   }
1077 }
1078
1079 void LoopInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1080   AU.setPreservesAll();
1081   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1082 }
1083
1084 void LoopInfoWrapperPass::print(raw_ostream &OS, const Module *) const {
1085   LI.print(OS);
1086 }
1087
1088 PreservedAnalyses LoopVerifierPass::run(Function &F,
1089                                         FunctionAnalysisManager &AM) {
1090   LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
1091   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1092   LI.verify(DT);
1093   return PreservedAnalyses::all();
1094 }
1095
1096 //===----------------------------------------------------------------------===//
1097 // LoopBlocksDFS implementation
1098 //
1099
1100 /// Traverse the loop blocks and store the DFS result.
1101 /// Useful for clients that just want the final DFS result and don't need to
1102 /// visit blocks during the initial traversal.
1103 void LoopBlocksDFS::perform(LoopInfo *LI) {
1104   LoopBlocksTraversal Traversal(*this, LI);
1105   for (LoopBlocksTraversal::POTIterator POI = Traversal.begin(),
1106                                         POE = Traversal.end();
1107        POI != POE; ++POI)
1108     ;
1109 }