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/InitializePasses.h" 38 #include "llvm/Support/CommandLine.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/raw_ostream.h" 41 #include <algorithm> 42 using namespace llvm; 43 44 // Explicitly instantiate methods in LoopInfoImpl.h for IR-level Loops. 45 template class llvm::LoopBase<BasicBlock, Loop>; 46 template class llvm::LoopInfoBase<BasicBlock, Loop>; 47 48 // Always verify loopinfo if expensive checking is enabled. 49 #ifdef EXPENSIVE_CHECKS 50 bool llvm::VerifyLoopInfo = true; 51 #else 52 bool llvm::VerifyLoopInfo = false; 53 #endif 54 static cl::opt<bool, true> 55 VerifyLoopInfoX("verify-loop-info", cl::location(VerifyLoopInfo), 56 cl::Hidden, cl::desc("Verify loop info (time consuming)")); 57 58 //===----------------------------------------------------------------------===// 59 // Loop implementation 60 // 61 62 bool Loop::isLoopInvariant(const Value *V) const { 63 if (const Instruction *I = dyn_cast<Instruction>(V)) 64 return !contains(I); 65 return true; // All non-instructions are loop invariant 66 } 67 68 bool Loop::hasLoopInvariantOperands(const Instruction *I) const { 69 return all_of(I->operands(), [this](Value *V) { return isLoopInvariant(V); }); 70 } 71 72 bool Loop::makeLoopInvariant(Value *V, bool &Changed, Instruction *InsertPt, 73 MemorySSAUpdater *MSSAU) const { 74 if (Instruction *I = dyn_cast<Instruction>(V)) 75 return makeLoopInvariant(I, Changed, InsertPt, MSSAU); 76 return true; // All non-instructions are loop-invariant. 77 } 78 79 bool Loop::makeLoopInvariant(Instruction *I, bool &Changed, 80 Instruction *InsertPt, 81 MemorySSAUpdater *MSSAU) const { 82 // Test if the value is already loop-invariant. 83 if (isLoopInvariant(I)) 84 return true; 85 if (!isSafeToSpeculativelyExecute(I)) 86 return false; 87 if (I->mayReadFromMemory()) 88 return false; 89 // EH block instructions are immobile. 90 if (I->isEHPad()) 91 return false; 92 // Determine the insertion point, unless one was given. 93 if (!InsertPt) { 94 BasicBlock *Preheader = getLoopPreheader(); 95 // Without a preheader, hoisting is not feasible. 96 if (!Preheader) 97 return false; 98 InsertPt = Preheader->getTerminator(); 99 } 100 // Don't hoist instructions with loop-variant operands. 101 for (Value *Operand : I->operands()) 102 if (!makeLoopInvariant(Operand, Changed, InsertPt, MSSAU)) 103 return false; 104 105 // Hoist. 106 I->moveBefore(InsertPt); 107 if (MSSAU) 108 if (auto *MUD = MSSAU->getMemorySSA()->getMemoryAccess(I)) 109 MSSAU->moveToPlace(MUD, InsertPt->getParent(), 110 MemorySSA::BeforeTerminator); 111 112 // There is possibility of hoisting this instruction above some arbitrary 113 // condition. Any metadata defined on it can be control dependent on this 114 // condition. Conservatively strip it here so that we don't give any wrong 115 // information to the optimizer. 116 I->dropUnknownNonDebugMetadata(); 117 118 Changed = true; 119 return true; 120 } 121 122 bool Loop::getIncomingAndBackEdge(BasicBlock *&Incoming, 123 BasicBlock *&Backedge) const { 124 BasicBlock *H = getHeader(); 125 126 Incoming = nullptr; 127 Backedge = nullptr; 128 pred_iterator PI = pred_begin(H); 129 assert(PI != pred_end(H) && "Loop must have at least one backedge!"); 130 Backedge = *PI++; 131 if (PI == pred_end(H)) 132 return false; // dead loop 133 Incoming = *PI++; 134 if (PI != pred_end(H)) 135 return false; // multiple backedges? 136 137 if (contains(Incoming)) { 138 if (contains(Backedge)) 139 return false; 140 std::swap(Incoming, Backedge); 141 } else if (!contains(Backedge)) 142 return false; 143 144 assert(Incoming && Backedge && "expected non-null incoming and backedges"); 145 return true; 146 } 147 148 PHINode *Loop::getCanonicalInductionVariable() const { 149 BasicBlock *H = getHeader(); 150 151 BasicBlock *Incoming = nullptr, *Backedge = nullptr; 152 if (!getIncomingAndBackEdge(Incoming, Backedge)) 153 return nullptr; 154 155 // Loop over all of the PHI nodes, looking for a canonical indvar. 156 for (BasicBlock::iterator I = H->begin(); isa<PHINode>(I); ++I) { 157 PHINode *PN = cast<PHINode>(I); 158 if (ConstantInt *CI = 159 dyn_cast<ConstantInt>(PN->getIncomingValueForBlock(Incoming))) 160 if (CI->isZero()) 161 if (Instruction *Inc = 162 dyn_cast<Instruction>(PN->getIncomingValueForBlock(Backedge))) 163 if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(0) == PN) 164 if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1))) 165 if (CI->isOne()) 166 return PN; 167 } 168 return nullptr; 169 } 170 171 /// Get the latch condition instruction. 172 static ICmpInst *getLatchCmpInst(const Loop &L) { 173 if (BasicBlock *Latch = L.getLoopLatch()) 174 if (BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator())) 175 if (BI->isConditional()) 176 return dyn_cast<ICmpInst>(BI->getCondition()); 177 178 return nullptr; 179 } 180 181 /// Return the final value of the loop induction variable if found. 182 static Value *findFinalIVValue(const Loop &L, const PHINode &IndVar, 183 const Instruction &StepInst) { 184 ICmpInst *LatchCmpInst = getLatchCmpInst(L); 185 if (!LatchCmpInst) 186 return nullptr; 187 188 Value *Op0 = LatchCmpInst->getOperand(0); 189 Value *Op1 = LatchCmpInst->getOperand(1); 190 if (Op0 == &IndVar || Op0 == &StepInst) 191 return Op1; 192 193 if (Op1 == &IndVar || Op1 == &StepInst) 194 return Op0; 195 196 return nullptr; 197 } 198 199 Optional<Loop::LoopBounds> Loop::LoopBounds::getBounds(const Loop &L, 200 PHINode &IndVar, 201 ScalarEvolution &SE) { 202 InductionDescriptor IndDesc; 203 if (!InductionDescriptor::isInductionPHI(&IndVar, &L, &SE, IndDesc)) 204 return None; 205 206 Value *InitialIVValue = IndDesc.getStartValue(); 207 Instruction *StepInst = IndDesc.getInductionBinOp(); 208 if (!InitialIVValue || !StepInst) 209 return None; 210 211 const SCEV *Step = IndDesc.getStep(); 212 Value *StepInstOp1 = StepInst->getOperand(1); 213 Value *StepInstOp0 = StepInst->getOperand(0); 214 Value *StepValue = nullptr; 215 if (SE.getSCEV(StepInstOp1) == Step) 216 StepValue = StepInstOp1; 217 else if (SE.getSCEV(StepInstOp0) == Step) 218 StepValue = StepInstOp0; 219 220 Value *FinalIVValue = findFinalIVValue(L, IndVar, *StepInst); 221 if (!FinalIVValue) 222 return None; 223 224 return LoopBounds(L, *InitialIVValue, *StepInst, StepValue, *FinalIVValue, 225 SE); 226 } 227 228 using Direction = Loop::LoopBounds::Direction; 229 230 ICmpInst::Predicate Loop::LoopBounds::getCanonicalPredicate() const { 231 BasicBlock *Latch = L.getLoopLatch(); 232 assert(Latch && "Expecting valid latch"); 233 234 BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator()); 235 assert(BI && BI->isConditional() && "Expecting conditional latch branch"); 236 237 ICmpInst *LatchCmpInst = dyn_cast<ICmpInst>(BI->getCondition()); 238 assert(LatchCmpInst && 239 "Expecting the latch compare instruction to be a CmpInst"); 240 241 // Need to inverse the predicate when first successor is not the loop 242 // header 243 ICmpInst::Predicate Pred = (BI->getSuccessor(0) == L.getHeader()) 244 ? LatchCmpInst->getPredicate() 245 : LatchCmpInst->getInversePredicate(); 246 247 if (LatchCmpInst->getOperand(0) == &getFinalIVValue()) 248 Pred = ICmpInst::getSwappedPredicate(Pred); 249 250 // Need to flip strictness of the predicate when the latch compare instruction 251 // is not using StepInst 252 if (LatchCmpInst->getOperand(0) == &getStepInst() || 253 LatchCmpInst->getOperand(1) == &getStepInst()) 254 return Pred; 255 256 // Cannot flip strictness of NE and EQ 257 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 258 return ICmpInst::getFlippedStrictnessPredicate(Pred); 259 260 Direction D = getDirection(); 261 if (D == Direction::Increasing) 262 return ICmpInst::ICMP_SLT; 263 264 if (D == Direction::Decreasing) 265 return ICmpInst::ICMP_SGT; 266 267 // If cannot determine the direction, then unable to find the canonical 268 // predicate 269 return ICmpInst::BAD_ICMP_PREDICATE; 270 } 271 272 Direction Loop::LoopBounds::getDirection() const { 273 if (const SCEVAddRecExpr *StepAddRecExpr = 274 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&getStepInst()))) 275 if (const SCEV *StepRecur = StepAddRecExpr->getStepRecurrence(SE)) { 276 if (SE.isKnownPositive(StepRecur)) 277 return Direction::Increasing; 278 if (SE.isKnownNegative(StepRecur)) 279 return Direction::Decreasing; 280 } 281 282 return Direction::Unknown; 283 } 284 285 Optional<Loop::LoopBounds> Loop::getBounds(ScalarEvolution &SE) const { 286 if (PHINode *IndVar = getInductionVariable(SE)) 287 return LoopBounds::getBounds(*this, *IndVar, SE); 288 289 return None; 290 } 291 292 PHINode *Loop::getInductionVariable(ScalarEvolution &SE) const { 293 if (!isLoopSimplifyForm()) 294 return nullptr; 295 296 BasicBlock *Header = getHeader(); 297 assert(Header && "Expected a valid loop header"); 298 ICmpInst *CmpInst = getLatchCmpInst(*this); 299 if (!CmpInst) 300 return nullptr; 301 302 Instruction *LatchCmpOp0 = dyn_cast<Instruction>(CmpInst->getOperand(0)); 303 Instruction *LatchCmpOp1 = dyn_cast<Instruction>(CmpInst->getOperand(1)); 304 305 for (PHINode &IndVar : Header->phis()) { 306 InductionDescriptor IndDesc; 307 if (!InductionDescriptor::isInductionPHI(&IndVar, this, &SE, IndDesc)) 308 continue; 309 310 Instruction *StepInst = IndDesc.getInductionBinOp(); 311 312 // case 1: 313 // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}] 314 // StepInst = IndVar + step 315 // cmp = StepInst < FinalValue 316 if (StepInst == LatchCmpOp0 || StepInst == LatchCmpOp1) 317 return &IndVar; 318 319 // case 2: 320 // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}] 321 // StepInst = IndVar + step 322 // cmp = IndVar < FinalValue 323 if (&IndVar == LatchCmpOp0 || &IndVar == LatchCmpOp1) 324 return &IndVar; 325 } 326 327 return nullptr; 328 } 329 330 bool Loop::getInductionDescriptor(ScalarEvolution &SE, 331 InductionDescriptor &IndDesc) const { 332 if (PHINode *IndVar = getInductionVariable(SE)) 333 return InductionDescriptor::isInductionPHI(IndVar, this, &SE, IndDesc); 334 335 return false; 336 } 337 338 bool Loop::isAuxiliaryInductionVariable(PHINode &AuxIndVar, 339 ScalarEvolution &SE) const { 340 // Located in the loop header 341 BasicBlock *Header = getHeader(); 342 if (AuxIndVar.getParent() != Header) 343 return false; 344 345 // No uses outside of the loop 346 for (User *U : AuxIndVar.users()) 347 if (const Instruction *I = dyn_cast<Instruction>(U)) 348 if (!contains(I)) 349 return false; 350 351 InductionDescriptor IndDesc; 352 if (!InductionDescriptor::isInductionPHI(&AuxIndVar, this, &SE, IndDesc)) 353 return false; 354 355 // The step instruction opcode should be add or sub. 356 if (IndDesc.getInductionOpcode() != Instruction::Add && 357 IndDesc.getInductionOpcode() != Instruction::Sub) 358 return false; 359 360 // Incremented by a loop invariant step for each loop iteration 361 return SE.isLoopInvariant(IndDesc.getStep(), this); 362 } 363 364 BranchInst *Loop::getLoopGuardBranch() const { 365 if (!isLoopSimplifyForm()) 366 return nullptr; 367 368 BasicBlock *Preheader = getLoopPreheader(); 369 assert(Preheader && getLoopLatch() && 370 "Expecting a loop with valid preheader and latch"); 371 372 // Loop should be in rotate form. 373 if (!isRotatedForm()) 374 return nullptr; 375 376 // Disallow loops with more than one unique exit block, as we do not verify 377 // that GuardOtherSucc post dominates all exit blocks. 378 BasicBlock *ExitFromLatch = getUniqueExitBlock(); 379 if (!ExitFromLatch) 380 return nullptr; 381 382 BasicBlock *ExitFromLatchSucc = ExitFromLatch->getUniqueSuccessor(); 383 if (!ExitFromLatchSucc) 384 return nullptr; 385 386 BasicBlock *GuardBB = Preheader->getUniquePredecessor(); 387 if (!GuardBB) 388 return nullptr; 389 390 assert(GuardBB->getTerminator() && "Expecting valid guard terminator"); 391 392 BranchInst *GuardBI = dyn_cast<BranchInst>(GuardBB->getTerminator()); 393 if (!GuardBI || GuardBI->isUnconditional()) 394 return nullptr; 395 396 BasicBlock *GuardOtherSucc = (GuardBI->getSuccessor(0) == Preheader) 397 ? GuardBI->getSuccessor(1) 398 : GuardBI->getSuccessor(0); 399 return (GuardOtherSucc == ExitFromLatchSucc) ? GuardBI : nullptr; 400 } 401 402 bool Loop::isCanonical(ScalarEvolution &SE) const { 403 InductionDescriptor IndDesc; 404 if (!getInductionDescriptor(SE, IndDesc)) 405 return false; 406 407 ConstantInt *Init = dyn_cast_or_null<ConstantInt>(IndDesc.getStartValue()); 408 if (!Init || !Init->isZero()) 409 return false; 410 411 if (IndDesc.getInductionOpcode() != Instruction::Add) 412 return false; 413 414 ConstantInt *Step = IndDesc.getConstIntStepValue(); 415 if (!Step || !Step->isOne()) 416 return false; 417 418 return true; 419 } 420 421 // Check that 'BB' doesn't have any uses outside of the 'L' 422 static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB, 423 const DominatorTree &DT) { 424 for (const Instruction &I : BB) { 425 // Tokens can't be used in PHI nodes and live-out tokens prevent loop 426 // optimizations, so for the purposes of considered LCSSA form, we 427 // can ignore them. 428 if (I.getType()->isTokenTy()) 429 continue; 430 431 for (const Use &U : I.uses()) { 432 const Instruction *UI = cast<Instruction>(U.getUser()); 433 const BasicBlock *UserBB = UI->getParent(); 434 435 // For practical purposes, we consider that the use in a PHI 436 // occurs in the respective predecessor block. For more info, 437 // see the `phi` doc in LangRef and the LCSSA doc. 438 if (const PHINode *P = dyn_cast<PHINode>(UI)) 439 UserBB = P->getIncomingBlock(U); 440 441 // Check the current block, as a fast-path, before checking whether 442 // the use is anywhere in the loop. Most values are used in the same 443 // block they are defined in. Also, blocks not reachable from the 444 // entry are special; uses in them don't need to go through PHIs. 445 if (UserBB != &BB && !L.contains(UserBB) && 446 DT.isReachableFromEntry(UserBB)) 447 return false; 448 } 449 } 450 return true; 451 } 452 453 bool Loop::isLCSSAForm(const DominatorTree &DT) const { 454 // For each block we check that it doesn't have any uses outside of this loop. 455 return all_of(this->blocks(), [&](const BasicBlock *BB) { 456 return isBlockInLCSSAForm(*this, *BB, DT); 457 }); 458 } 459 460 bool Loop::isRecursivelyLCSSAForm(const DominatorTree &DT, 461 const LoopInfo &LI) const { 462 // For each block we check that it doesn't have any uses outside of its 463 // innermost loop. This process will transitively guarantee that the current 464 // loop and all of the nested loops are in LCSSA form. 465 return all_of(this->blocks(), [&](const BasicBlock *BB) { 466 return isBlockInLCSSAForm(*LI.getLoopFor(BB), *BB, DT); 467 }); 468 } 469 470 bool Loop::isLoopSimplifyForm() const { 471 // Normal-form loops have a preheader, a single backedge, and all of their 472 // exits have all their predecessors inside the loop. 473 return getLoopPreheader() && getLoopLatch() && hasDedicatedExits(); 474 } 475 476 // Routines that reform the loop CFG and split edges often fail on indirectbr. 477 bool Loop::isSafeToClone() const { 478 // Return false if any loop blocks contain indirectbrs, or there are any calls 479 // to noduplicate functions. 480 // FIXME: it should be ok to clone CallBrInst's if we correctly update the 481 // operand list to reflect the newly cloned labels. 482 for (BasicBlock *BB : this->blocks()) { 483 if (isa<IndirectBrInst>(BB->getTerminator()) || 484 isa<CallBrInst>(BB->getTerminator())) 485 return false; 486 487 for (Instruction &I : *BB) 488 if (auto *CB = dyn_cast<CallBase>(&I)) 489 if (CB->cannotDuplicate()) 490 return false; 491 } 492 return true; 493 } 494 495 MDNode *Loop::getLoopID() const { 496 MDNode *LoopID = nullptr; 497 498 // Go through the latch blocks and check the terminator for the metadata. 499 SmallVector<BasicBlock *, 4> LatchesBlocks; 500 getLoopLatches(LatchesBlocks); 501 for (BasicBlock *BB : LatchesBlocks) { 502 Instruction *TI = BB->getTerminator(); 503 MDNode *MD = TI->getMetadata(LLVMContext::MD_loop); 504 505 if (!MD) 506 return nullptr; 507 508 if (!LoopID) 509 LoopID = MD; 510 else if (MD != LoopID) 511 return nullptr; 512 } 513 if (!LoopID || LoopID->getNumOperands() == 0 || 514 LoopID->getOperand(0) != LoopID) 515 return nullptr; 516 return LoopID; 517 } 518 519 void Loop::setLoopID(MDNode *LoopID) const { 520 assert((!LoopID || LoopID->getNumOperands() > 0) && 521 "Loop ID needs at least one operand"); 522 assert((!LoopID || LoopID->getOperand(0) == LoopID) && 523 "Loop ID should refer to itself"); 524 525 SmallVector<BasicBlock *, 4> LoopLatches; 526 getLoopLatches(LoopLatches); 527 for (BasicBlock *BB : LoopLatches) 528 BB->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID); 529 } 530 531 void Loop::setLoopAlreadyUnrolled() { 532 LLVMContext &Context = getHeader()->getContext(); 533 534 MDNode *DisableUnrollMD = 535 MDNode::get(Context, MDString::get(Context, "llvm.loop.unroll.disable")); 536 MDNode *LoopID = getLoopID(); 537 MDNode *NewLoopID = makePostTransformationMetadata( 538 Context, LoopID, {"llvm.loop.unroll."}, {DisableUnrollMD}); 539 setLoopID(NewLoopID); 540 } 541 542 void Loop::setLoopMustProgress() { 543 LLVMContext &Context = getHeader()->getContext(); 544 545 MDNode *MustProgress = findOptionMDForLoop(this, "llvm.loop.mustprogress"); 546 547 if (MustProgress) 548 return; 549 550 MDNode *MustProgressMD = 551 MDNode::get(Context, MDString::get(Context, "llvm.loop.mustprogress")); 552 MDNode *LoopID = getLoopID(); 553 MDNode *NewLoopID = 554 makePostTransformationMetadata(Context, LoopID, {}, {MustProgressMD}); 555 setLoopID(NewLoopID); 556 } 557 558 bool Loop::isAnnotatedParallel() const { 559 MDNode *DesiredLoopIdMetadata = getLoopID(); 560 561 if (!DesiredLoopIdMetadata) 562 return false; 563 564 MDNode *ParallelAccesses = 565 findOptionMDForLoop(this, "llvm.loop.parallel_accesses"); 566 SmallPtrSet<MDNode *, 4> 567 ParallelAccessGroups; // For scalable 'contains' check. 568 if (ParallelAccesses) { 569 for (const MDOperand &MD : drop_begin(ParallelAccesses->operands(), 1)) { 570 MDNode *AccGroup = cast<MDNode>(MD.get()); 571 assert(isValidAsAccessGroup(AccGroup) && 572 "List item must be an access group"); 573 ParallelAccessGroups.insert(AccGroup); 574 } 575 } 576 577 // The loop branch contains the parallel loop metadata. In order to ensure 578 // that any parallel-loop-unaware optimization pass hasn't added loop-carried 579 // dependencies (thus converted the loop back to a sequential loop), check 580 // that all the memory instructions in the loop belong to an access group that 581 // is parallel to this loop. 582 for (BasicBlock *BB : this->blocks()) { 583 for (Instruction &I : *BB) { 584 if (!I.mayReadOrWriteMemory()) 585 continue; 586 587 if (MDNode *AccessGroup = I.getMetadata(LLVMContext::MD_access_group)) { 588 auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool { 589 if (AG->getNumOperands() == 0) { 590 assert(isValidAsAccessGroup(AG) && "Item must be an access group"); 591 return ParallelAccessGroups.count(AG); 592 } 593 594 for (const MDOperand &AccessListItem : AG->operands()) { 595 MDNode *AccGroup = cast<MDNode>(AccessListItem.get()); 596 assert(isValidAsAccessGroup(AccGroup) && 597 "List item must be an access group"); 598 if (ParallelAccessGroups.count(AccGroup)) 599 return true; 600 } 601 return false; 602 }; 603 604 if (ContainsAccessGroup(AccessGroup)) 605 continue; 606 } 607 608 // The memory instruction can refer to the loop identifier metadata 609 // directly or indirectly through another list metadata (in case of 610 // nested parallel loops). The loop identifier metadata refers to 611 // itself so we can check both cases with the same routine. 612 MDNode *LoopIdMD = 613 I.getMetadata(LLVMContext::MD_mem_parallel_loop_access); 614 615 if (!LoopIdMD) 616 return false; 617 618 bool LoopIdMDFound = false; 619 for (const MDOperand &MDOp : LoopIdMD->operands()) { 620 if (MDOp == DesiredLoopIdMetadata) { 621 LoopIdMDFound = true; 622 break; 623 } 624 } 625 626 if (!LoopIdMDFound) 627 return false; 628 } 629 } 630 return true; 631 } 632 633 DebugLoc Loop::getStartLoc() const { return getLocRange().getStart(); } 634 635 Loop::LocRange Loop::getLocRange() const { 636 // If we have a debug location in the loop ID, then use it. 637 if (MDNode *LoopID = getLoopID()) { 638 DebugLoc Start; 639 // We use the first DebugLoc in the header as the start location of the loop 640 // and if there is a second DebugLoc in the header we use it as end location 641 // of the loop. 642 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 643 if (DILocation *L = dyn_cast<DILocation>(LoopID->getOperand(i))) { 644 if (!Start) 645 Start = DebugLoc(L); 646 else 647 return LocRange(Start, DebugLoc(L)); 648 } 649 } 650 651 if (Start) 652 return LocRange(Start); 653 } 654 655 // Try the pre-header first. 656 if (BasicBlock *PHeadBB = getLoopPreheader()) 657 if (DebugLoc DL = PHeadBB->getTerminator()->getDebugLoc()) 658 return LocRange(DL); 659 660 // If we have no pre-header or there are no instructions with debug 661 // info in it, try the header. 662 if (BasicBlock *HeadBB = getHeader()) 663 return LocRange(HeadBB->getTerminator()->getDebugLoc()); 664 665 return LocRange(); 666 } 667 668 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 669 LLVM_DUMP_METHOD void Loop::dump() const { print(dbgs()); } 670 671 LLVM_DUMP_METHOD void Loop::dumpVerbose() const { 672 print(dbgs(), /*Depth=*/0, /*Verbose=*/true); 673 } 674 #endif 675 676 //===----------------------------------------------------------------------===// 677 // UnloopUpdater implementation 678 // 679 680 namespace { 681 /// Find the new parent loop for all blocks within the "unloop" whose last 682 /// backedges has just been removed. 683 class UnloopUpdater { 684 Loop &Unloop; 685 LoopInfo *LI; 686 687 LoopBlocksDFS DFS; 688 689 // Map unloop's immediate subloops to their nearest reachable parents. Nested 690 // loops within these subloops will not change parents. However, an immediate 691 // subloop's new parent will be the nearest loop reachable from either its own 692 // exits *or* any of its nested loop's exits. 693 DenseMap<Loop *, Loop *> SubloopParents; 694 695 // Flag the presence of an irreducible backedge whose destination is a block 696 // directly contained by the original unloop. 697 bool FoundIB; 698 699 public: 700 UnloopUpdater(Loop *UL, LoopInfo *LInfo) 701 : Unloop(*UL), LI(LInfo), DFS(UL), FoundIB(false) {} 702 703 void updateBlockParents(); 704 705 void removeBlocksFromAncestors(); 706 707 void updateSubloopParents(); 708 709 protected: 710 Loop *getNearestLoop(BasicBlock *BB, Loop *BBLoop); 711 }; 712 } // end anonymous namespace 713 714 /// Update the parent loop for all blocks that are directly contained within the 715 /// original "unloop". 716 void UnloopUpdater::updateBlockParents() { 717 if (Unloop.getNumBlocks()) { 718 // Perform a post order CFG traversal of all blocks within this loop, 719 // propagating the nearest loop from successors to predecessors. 720 LoopBlocksTraversal Traversal(DFS, LI); 721 for (BasicBlock *POI : Traversal) { 722 723 Loop *L = LI->getLoopFor(POI); 724 Loop *NL = getNearestLoop(POI, L); 725 726 if (NL != L) { 727 // For reducible loops, NL is now an ancestor of Unloop. 728 assert((NL != &Unloop && (!NL || NL->contains(&Unloop))) && 729 "uninitialized successor"); 730 LI->changeLoopFor(POI, NL); 731 } else { 732 // Or the current block is part of a subloop, in which case its parent 733 // is unchanged. 734 assert((FoundIB || Unloop.contains(L)) && "uninitialized successor"); 735 } 736 } 737 } 738 // Each irreducible loop within the unloop induces a round of iteration using 739 // the DFS result cached by Traversal. 740 bool Changed = FoundIB; 741 for (unsigned NIters = 0; Changed; ++NIters) { 742 assert(NIters < Unloop.getNumBlocks() && "runaway iterative algorithm"); 743 744 // Iterate over the postorder list of blocks, propagating the nearest loop 745 // from successors to predecessors as before. 746 Changed = false; 747 for (LoopBlocksDFS::POIterator POI = DFS.beginPostorder(), 748 POE = DFS.endPostorder(); 749 POI != POE; ++POI) { 750 751 Loop *L = LI->getLoopFor(*POI); 752 Loop *NL = getNearestLoop(*POI, L); 753 if (NL != L) { 754 assert(NL != &Unloop && (!NL || NL->contains(&Unloop)) && 755 "uninitialized successor"); 756 LI->changeLoopFor(*POI, NL); 757 Changed = true; 758 } 759 } 760 } 761 } 762 763 /// Remove unloop's blocks from all ancestors below their new parents. 764 void UnloopUpdater::removeBlocksFromAncestors() { 765 // Remove all unloop's blocks (including those in nested subloops) from 766 // ancestors below the new parent loop. 767 for (Loop::block_iterator BI = Unloop.block_begin(), BE = Unloop.block_end(); 768 BI != BE; ++BI) { 769 Loop *OuterParent = LI->getLoopFor(*BI); 770 if (Unloop.contains(OuterParent)) { 771 while (OuterParent->getParentLoop() != &Unloop) 772 OuterParent = OuterParent->getParentLoop(); 773 OuterParent = SubloopParents[OuterParent]; 774 } 775 // Remove blocks from former Ancestors except Unloop itself which will be 776 // deleted. 777 for (Loop *OldParent = Unloop.getParentLoop(); OldParent != OuterParent; 778 OldParent = OldParent->getParentLoop()) { 779 assert(OldParent && "new loop is not an ancestor of the original"); 780 OldParent->removeBlockFromLoop(*BI); 781 } 782 } 783 } 784 785 /// Update the parent loop for all subloops directly nested within unloop. 786 void UnloopUpdater::updateSubloopParents() { 787 while (!Unloop.isInnermost()) { 788 Loop *Subloop = *std::prev(Unloop.end()); 789 Unloop.removeChildLoop(std::prev(Unloop.end())); 790 791 assert(SubloopParents.count(Subloop) && "DFS failed to visit subloop"); 792 if (Loop *Parent = SubloopParents[Subloop]) 793 Parent->addChildLoop(Subloop); 794 else 795 LI->addTopLevelLoop(Subloop); 796 } 797 } 798 799 /// Return the nearest parent loop among this block's successors. If a successor 800 /// is a subloop header, consider its parent to be the nearest parent of the 801 /// subloop's exits. 802 /// 803 /// For subloop blocks, simply update SubloopParents and return NULL. 804 Loop *UnloopUpdater::getNearestLoop(BasicBlock *BB, Loop *BBLoop) { 805 806 // Initially for blocks directly contained by Unloop, NearLoop == Unloop and 807 // is considered uninitialized. 808 Loop *NearLoop = BBLoop; 809 810 Loop *Subloop = nullptr; 811 if (NearLoop != &Unloop && Unloop.contains(NearLoop)) { 812 Subloop = NearLoop; 813 // Find the subloop ancestor that is directly contained within Unloop. 814 while (Subloop->getParentLoop() != &Unloop) { 815 Subloop = Subloop->getParentLoop(); 816 assert(Subloop && "subloop is not an ancestor of the original loop"); 817 } 818 // Get the current nearest parent of the Subloop exits, initially Unloop. 819 NearLoop = SubloopParents.insert({Subloop, &Unloop}).first->second; 820 } 821 822 succ_iterator I = succ_begin(BB), E = succ_end(BB); 823 if (I == E) { 824 assert(!Subloop && "subloop blocks must have a successor"); 825 NearLoop = nullptr; // unloop blocks may now exit the function. 826 } 827 for (; I != E; ++I) { 828 if (*I == BB) 829 continue; // self loops are uninteresting 830 831 Loop *L = LI->getLoopFor(*I); 832 if (L == &Unloop) { 833 // This successor has not been processed. This path must lead to an 834 // irreducible backedge. 835 assert((FoundIB || !DFS.hasPostorder(*I)) && "should have seen IB"); 836 FoundIB = true; 837 } 838 if (L != &Unloop && Unloop.contains(L)) { 839 // Successor is in a subloop. 840 if (Subloop) 841 continue; // Branching within subloops. Ignore it. 842 843 // BB branches from the original into a subloop header. 844 assert(L->getParentLoop() == &Unloop && "cannot skip into nested loops"); 845 846 // Get the current nearest parent of the Subloop's exits. 847 L = SubloopParents[L]; 848 // L could be Unloop if the only exit was an irreducible backedge. 849 } 850 if (L == &Unloop) { 851 continue; 852 } 853 // Handle critical edges from Unloop into a sibling loop. 854 if (L && !L->contains(&Unloop)) { 855 L = L->getParentLoop(); 856 } 857 // Remember the nearest parent loop among successors or subloop exits. 858 if (NearLoop == &Unloop || !NearLoop || NearLoop->contains(L)) 859 NearLoop = L; 860 } 861 if (Subloop) { 862 SubloopParents[Subloop] = NearLoop; 863 return BBLoop; 864 } 865 return NearLoop; 866 } 867 868 LoopInfo::LoopInfo(const DomTreeBase<BasicBlock> &DomTree) { analyze(DomTree); } 869 870 bool LoopInfo::invalidate(Function &F, const PreservedAnalyses &PA, 871 FunctionAnalysisManager::Invalidator &) { 872 // Check whether the analysis, all analyses on functions, or the function's 873 // CFG have been preserved. 874 auto PAC = PA.getChecker<LoopAnalysis>(); 875 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() || 876 PAC.preservedSet<CFGAnalyses>()); 877 } 878 879 void LoopInfo::erase(Loop *Unloop) { 880 assert(!Unloop->isInvalid() && "Loop has already been erased!"); 881 882 auto InvalidateOnExit = make_scope_exit([&]() { destroy(Unloop); }); 883 884 // First handle the special case of no parent loop to simplify the algorithm. 885 if (Unloop->isOutermost()) { 886 // Since BBLoop had no parent, Unloop blocks are no longer in a loop. 887 for (Loop::block_iterator I = Unloop->block_begin(), 888 E = Unloop->block_end(); 889 I != E; ++I) { 890 891 // Don't reparent blocks in subloops. 892 if (getLoopFor(*I) != Unloop) 893 continue; 894 895 // Blocks no longer have a parent but are still referenced by Unloop until 896 // the Unloop object is deleted. 897 changeLoopFor(*I, nullptr); 898 } 899 900 // Remove the loop from the top-level LoopInfo object. 901 for (iterator I = begin();; ++I) { 902 assert(I != end() && "Couldn't find loop"); 903 if (*I == Unloop) { 904 removeLoop(I); 905 break; 906 } 907 } 908 909 // Move all of the subloops to the top-level. 910 while (!Unloop->isInnermost()) 911 addTopLevelLoop(Unloop->removeChildLoop(std::prev(Unloop->end()))); 912 913 return; 914 } 915 916 // Update the parent loop for all blocks within the loop. Blocks within 917 // subloops will not change parents. 918 UnloopUpdater Updater(Unloop, this); 919 Updater.updateBlockParents(); 920 921 // Remove blocks from former ancestor loops. 922 Updater.removeBlocksFromAncestors(); 923 924 // Add direct subloops as children in their new parent loop. 925 Updater.updateSubloopParents(); 926 927 // Remove unloop from its parent loop. 928 Loop *ParentLoop = Unloop->getParentLoop(); 929 for (Loop::iterator I = ParentLoop->begin();; ++I) { 930 assert(I != ParentLoop->end() && "Couldn't find loop"); 931 if (*I == Unloop) { 932 ParentLoop->removeChildLoop(I); 933 break; 934 } 935 } 936 } 937 938 AnalysisKey LoopAnalysis::Key; 939 940 LoopInfo LoopAnalysis::run(Function &F, FunctionAnalysisManager &AM) { 941 // FIXME: Currently we create a LoopInfo from scratch for every function. 942 // This may prove to be too wasteful due to deallocating and re-allocating 943 // memory each time for the underlying map and vector datastructures. At some 944 // point it may prove worthwhile to use a freelist and recycle LoopInfo 945 // objects. I don't want to add that kind of complexity until the scope of 946 // the problem is better understood. 947 LoopInfo LI; 948 LI.analyze(AM.getResult<DominatorTreeAnalysis>(F)); 949 return LI; 950 } 951 952 PreservedAnalyses LoopPrinterPass::run(Function &F, 953 FunctionAnalysisManager &AM) { 954 AM.getResult<LoopAnalysis>(F).print(OS); 955 return PreservedAnalyses::all(); 956 } 957 958 void llvm::printLoop(Loop &L, raw_ostream &OS, const std::string &Banner) { 959 960 if (forcePrintModuleIR()) { 961 // handling -print-module-scope 962 OS << Banner << " (loop: "; 963 L.getHeader()->printAsOperand(OS, false); 964 OS << ")\n"; 965 966 // printing whole module 967 OS << *L.getHeader()->getModule(); 968 return; 969 } 970 971 OS << Banner; 972 973 auto *PreHeader = L.getLoopPreheader(); 974 if (PreHeader) { 975 OS << "\n; Preheader:"; 976 PreHeader->print(OS); 977 OS << "\n; Loop:"; 978 } 979 980 for (auto *Block : L.blocks()) 981 if (Block) 982 Block->print(OS); 983 else 984 OS << "Printing <null> block"; 985 986 SmallVector<BasicBlock *, 8> ExitBlocks; 987 L.getExitBlocks(ExitBlocks); 988 if (!ExitBlocks.empty()) { 989 OS << "\n; Exit blocks"; 990 for (auto *Block : ExitBlocks) 991 if (Block) 992 Block->print(OS); 993 else 994 OS << "Printing <null> block"; 995 } 996 } 997 998 MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) { 999 // No loop metadata node, no loop properties. 1000 if (!LoopID) 1001 return nullptr; 1002 1003 // First operand should refer to the metadata node itself, for legacy reasons. 1004 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 1005 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 1006 1007 // Iterate over the metdata node operands and look for MDString metadata. 1008 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 1009 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 1010 if (!MD || MD->getNumOperands() < 1) 1011 continue; 1012 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 1013 if (!S) 1014 continue; 1015 // Return the operand node if MDString holds expected metadata. 1016 if (Name.equals(S->getString())) 1017 return MD; 1018 } 1019 1020 // Loop property not found. 1021 return nullptr; 1022 } 1023 1024 MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) { 1025 return findOptionMDForLoopID(TheLoop->getLoopID(), Name); 1026 } 1027 1028 bool llvm::isValidAsAccessGroup(MDNode *Node) { 1029 return Node->getNumOperands() == 0 && Node->isDistinct(); 1030 } 1031 1032 MDNode *llvm::makePostTransformationMetadata(LLVMContext &Context, 1033 MDNode *OrigLoopID, 1034 ArrayRef<StringRef> RemovePrefixes, 1035 ArrayRef<MDNode *> AddAttrs) { 1036 // First remove any existing loop metadata related to this transformation. 1037 SmallVector<Metadata *, 4> MDs; 1038 1039 // Reserve first location for self reference to the LoopID metadata node. 1040 MDs.push_back(nullptr); 1041 1042 // Remove metadata for the transformation that has been applied or that became 1043 // outdated. 1044 if (OrigLoopID) { 1045 for (unsigned i = 1, ie = OrigLoopID->getNumOperands(); i < ie; ++i) { 1046 bool IsVectorMetadata = false; 1047 Metadata *Op = OrigLoopID->getOperand(i); 1048 if (MDNode *MD = dyn_cast<MDNode>(Op)) { 1049 const MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 1050 if (S) 1051 IsVectorMetadata = 1052 llvm::any_of(RemovePrefixes, [S](StringRef Prefix) -> bool { 1053 return S->getString().startswith(Prefix); 1054 }); 1055 } 1056 if (!IsVectorMetadata) 1057 MDs.push_back(Op); 1058 } 1059 } 1060 1061 // Add metadata to avoid reapplying a transformation, such as 1062 // llvm.loop.unroll.disable and llvm.loop.isvectorized. 1063 MDs.append(AddAttrs.begin(), AddAttrs.end()); 1064 1065 MDNode *NewLoopID = MDNode::getDistinct(Context, MDs); 1066 // Replace the temporary node with a self-reference. 1067 NewLoopID->replaceOperandWith(0, NewLoopID); 1068 return NewLoopID; 1069 } 1070 1071 //===----------------------------------------------------------------------===// 1072 // LoopInfo implementation 1073 // 1074 1075 LoopInfoWrapperPass::LoopInfoWrapperPass() : FunctionPass(ID) { 1076 initializeLoopInfoWrapperPassPass(*PassRegistry::getPassRegistry()); 1077 } 1078 1079 char LoopInfoWrapperPass::ID = 0; 1080 INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops", "Natural Loop Information", 1081 true, true) 1082 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1083 INITIALIZE_PASS_END(LoopInfoWrapperPass, "loops", "Natural Loop Information", 1084 true, true) 1085 1086 bool LoopInfoWrapperPass::runOnFunction(Function &) { 1087 releaseMemory(); 1088 LI.analyze(getAnalysis<DominatorTreeWrapperPass>().getDomTree()); 1089 return false; 1090 } 1091 1092 void LoopInfoWrapperPass::verifyAnalysis() const { 1093 // LoopInfoWrapperPass is a FunctionPass, but verifying every loop in the 1094 // function each time verifyAnalysis is called is very expensive. The 1095 // -verify-loop-info option can enable this. In order to perform some 1096 // checking by default, LoopPass has been taught to call verifyLoop manually 1097 // during loop pass sequences. 1098 if (VerifyLoopInfo) { 1099 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1100 LI.verify(DT); 1101 } 1102 } 1103 1104 void LoopInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 1105 AU.setPreservesAll(); 1106 AU.addRequiredTransitive<DominatorTreeWrapperPass>(); 1107 } 1108 1109 void LoopInfoWrapperPass::print(raw_ostream &OS, const Module *) const { 1110 LI.print(OS); 1111 } 1112 1113 PreservedAnalyses LoopVerifierPass::run(Function &F, 1114 FunctionAnalysisManager &AM) { 1115 LoopInfo &LI = AM.getResult<LoopAnalysis>(F); 1116 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1117 LI.verify(DT); 1118 return PreservedAnalyses::all(); 1119 } 1120 1121 //===----------------------------------------------------------------------===// 1122 // LoopBlocksDFS implementation 1123 // 1124 1125 /// Traverse the loop blocks and store the DFS result. 1126 /// Useful for clients that just want the final DFS result and don't need to 1127 /// visit blocks during the initial traversal. 1128 void LoopBlocksDFS::perform(LoopInfo *LI) { 1129 LoopBlocksTraversal Traversal(*this, LI); 1130 for (LoopBlocksTraversal::POTIterator POI = Traversal.begin(), 1131 POE = Traversal.end(); 1132 POI != POE; ++POI) 1133 ; 1134 } 1135