1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===// 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 implements some loop unrolling utilities. It does not define any 10 // actual pass or policy, but provides a single function to perform loop 11 // unrolling. 12 // 13 // The process of unrolling can produce extraneous basic blocks linked with 14 // unconditional branches. This will be corrected in the future. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "llvm/ADT/ArrayRef.h" 19 #include "llvm/ADT/DenseMap.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/ADT/StringRef.h" 25 #include "llvm/ADT/Twine.h" 26 #include "llvm/ADT/ilist_iterator.h" 27 #include "llvm/Analysis/AssumptionCache.h" 28 #include "llvm/Analysis/DomTreeUpdater.h" 29 #include "llvm/Analysis/InstructionSimplify.h" 30 #include "llvm/Analysis/LoopInfo.h" 31 #include "llvm/Analysis/LoopIterator.h" 32 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 33 #include "llvm/Analysis/ScalarEvolution.h" 34 #include "llvm/IR/BasicBlock.h" 35 #include "llvm/IR/CFG.h" 36 #include "llvm/IR/Constants.h" 37 #include "llvm/IR/DebugInfoMetadata.h" 38 #include "llvm/IR/DebugLoc.h" 39 #include "llvm/IR/DiagnosticInfo.h" 40 #include "llvm/IR/Dominators.h" 41 #include "llvm/IR/Function.h" 42 #include "llvm/IR/Instruction.h" 43 #include "llvm/IR/Instructions.h" 44 #include "llvm/IR/IntrinsicInst.h" 45 #include "llvm/IR/Metadata.h" 46 #include "llvm/IR/Module.h" 47 #include "llvm/IR/PatternMatch.h" 48 #include "llvm/IR/Use.h" 49 #include "llvm/IR/User.h" 50 #include "llvm/IR/ValueHandle.h" 51 #include "llvm/IR/ValueMap.h" 52 #include "llvm/Support/Casting.h" 53 #include "llvm/Support/CommandLine.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/GenericDomTree.h" 56 #include "llvm/Support/MathExtras.h" 57 #include "llvm/Support/raw_ostream.h" 58 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 59 #include "llvm/Transforms/Utils/Cloning.h" 60 #include "llvm/Transforms/Utils/Local.h" 61 #include "llvm/Transforms/Utils/LoopSimplify.h" 62 #include "llvm/Transforms/Utils/LoopUtils.h" 63 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 64 #include "llvm/Transforms/Utils/UnrollLoop.h" 65 #include "llvm/Transforms/Utils/ValueMapper.h" 66 #include <algorithm> 67 #include <assert.h> 68 #include <numeric> 69 #include <type_traits> 70 #include <vector> 71 72 namespace llvm { 73 class DataLayout; 74 class Value; 75 } // namespace llvm 76 77 using namespace llvm; 78 79 #define DEBUG_TYPE "loop-unroll" 80 81 // TODO: Should these be here or in LoopUnroll? 82 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled"); 83 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)"); 84 STATISTIC(NumUnrolledNotLatch, "Number of loops unrolled without a conditional " 85 "latch (completely or otherwise)"); 86 87 static cl::opt<bool> 88 UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden, 89 cl::desc("Allow runtime unrolled loops to be unrolled " 90 "with epilog instead of prolog.")); 91 92 static cl::opt<bool> 93 UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden, 94 cl::desc("Verify domtree after unrolling"), 95 #ifdef EXPENSIVE_CHECKS 96 cl::init(true) 97 #else 98 cl::init(false) 99 #endif 100 ); 101 102 static cl::opt<bool> 103 UnrollVerifyLoopInfo("unroll-verify-loopinfo", cl::Hidden, 104 cl::desc("Verify loopinfo after unrolling"), 105 #ifdef EXPENSIVE_CHECKS 106 cl::init(true) 107 #else 108 cl::init(false) 109 #endif 110 ); 111 112 113 /// Check if unrolling created a situation where we need to insert phi nodes to 114 /// preserve LCSSA form. 115 /// \param Blocks is a vector of basic blocks representing unrolled loop. 116 /// \param L is the outer loop. 117 /// It's possible that some of the blocks are in L, and some are not. In this 118 /// case, if there is a use is outside L, and definition is inside L, we need to 119 /// insert a phi-node, otherwise LCSSA will be broken. 120 /// The function is just a helper function for llvm::UnrollLoop that returns 121 /// true if this situation occurs, indicating that LCSSA needs to be fixed. 122 static bool needToInsertPhisForLCSSA(Loop *L, 123 const std::vector<BasicBlock *> &Blocks, 124 LoopInfo *LI) { 125 for (BasicBlock *BB : Blocks) { 126 if (LI->getLoopFor(BB) == L) 127 continue; 128 for (Instruction &I : *BB) { 129 for (Use &U : I.operands()) { 130 if (const auto *Def = dyn_cast<Instruction>(U)) { 131 Loop *DefLoop = LI->getLoopFor(Def->getParent()); 132 if (!DefLoop) 133 continue; 134 if (DefLoop->contains(L)) 135 return true; 136 } 137 } 138 } 139 } 140 return false; 141 } 142 143 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary 144 /// and adds a mapping from the original loop to the new loop to NewLoops. 145 /// Returns nullptr if no new loop was created and a pointer to the 146 /// original loop OriginalBB was part of otherwise. 147 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB, 148 BasicBlock *ClonedBB, LoopInfo *LI, 149 NewLoopsMap &NewLoops) { 150 // Figure out which loop New is in. 151 const Loop *OldLoop = LI->getLoopFor(OriginalBB); 152 assert(OldLoop && "Should (at least) be in the loop being unrolled!"); 153 154 Loop *&NewLoop = NewLoops[OldLoop]; 155 if (!NewLoop) { 156 // Found a new sub-loop. 157 assert(OriginalBB == OldLoop->getHeader() && 158 "Header should be first in RPO"); 159 160 NewLoop = LI->AllocateLoop(); 161 Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop()); 162 163 if (NewLoopParent) 164 NewLoopParent->addChildLoop(NewLoop); 165 else 166 LI->addTopLevelLoop(NewLoop); 167 168 NewLoop->addBasicBlockToLoop(ClonedBB, *LI); 169 return OldLoop; 170 } else { 171 NewLoop->addBasicBlockToLoop(ClonedBB, *LI); 172 return nullptr; 173 } 174 } 175 176 /// The function chooses which type of unroll (epilog or prolog) is more 177 /// profitabale. 178 /// Epilog unroll is more profitable when there is PHI that starts from 179 /// constant. In this case epilog will leave PHI start from constant, 180 /// but prolog will convert it to non-constant. 181 /// 182 /// loop: 183 /// PN = PHI [I, Latch], [CI, PreHeader] 184 /// I = foo(PN) 185 /// ... 186 /// 187 /// Epilog unroll case. 188 /// loop: 189 /// PN = PHI [I2, Latch], [CI, PreHeader] 190 /// I1 = foo(PN) 191 /// I2 = foo(I1) 192 /// ... 193 /// Prolog unroll case. 194 /// NewPN = PHI [PrologI, Prolog], [CI, PreHeader] 195 /// loop: 196 /// PN = PHI [I2, Latch], [NewPN, PreHeader] 197 /// I1 = foo(PN) 198 /// I2 = foo(I1) 199 /// ... 200 /// 201 static bool isEpilogProfitable(Loop *L) { 202 BasicBlock *PreHeader = L->getLoopPreheader(); 203 BasicBlock *Header = L->getHeader(); 204 assert(PreHeader && Header); 205 for (const PHINode &PN : Header->phis()) { 206 if (isa<ConstantInt>(PN.getIncomingValueForBlock(PreHeader))) 207 return true; 208 } 209 return false; 210 } 211 212 /// Perform some cleanup and simplifications on loops after unrolling. It is 213 /// useful to simplify the IV's in the new loop, as well as do a quick 214 /// simplify/dce pass of the instructions. 215 void llvm::simplifyLoopAfterUnroll(Loop *L, bool SimplifyIVs, LoopInfo *LI, 216 ScalarEvolution *SE, DominatorTree *DT, 217 AssumptionCache *AC, 218 const TargetTransformInfo *TTI) { 219 using namespace llvm::PatternMatch; 220 221 // Simplify any new induction variables in the partially unrolled loop. 222 if (SE && SimplifyIVs) { 223 SmallVector<WeakTrackingVH, 16> DeadInsts; 224 simplifyLoopIVs(L, SE, DT, LI, TTI, DeadInsts); 225 226 // Aggressively clean up dead instructions that simplifyLoopIVs already 227 // identified. Any remaining should be cleaned up below. 228 while (!DeadInsts.empty()) { 229 Value *V = DeadInsts.pop_back_val(); 230 if (Instruction *Inst = dyn_cast_or_null<Instruction>(V)) 231 RecursivelyDeleteTriviallyDeadInstructions(Inst); 232 } 233 } 234 235 // At this point, the code is well formed. Perform constprop, instsimplify, 236 // and dce. 237 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 238 SmallVector<WeakTrackingVH, 16> DeadInsts; 239 for (BasicBlock *BB : L->getBlocks()) { 240 for (Instruction &Inst : llvm::make_early_inc_range(*BB)) { 241 if (Value *V = simplifyInstruction(&Inst, {DL, nullptr, DT, AC})) 242 if (LI->replacementPreservesLCSSAForm(&Inst, V)) 243 Inst.replaceAllUsesWith(V); 244 if (isInstructionTriviallyDead(&Inst)) 245 DeadInsts.emplace_back(&Inst); 246 247 // Fold ((add X, C1), C2) to (add X, C1+C2). This is very common in 248 // unrolled loops, and handling this early allows following code to 249 // identify the IV as a "simple recurrence" without first folding away 250 // a long chain of adds. 251 { 252 Value *X; 253 const APInt *C1, *C2; 254 if (match(&Inst, m_Add(m_Add(m_Value(X), m_APInt(C1)), m_APInt(C2)))) { 255 auto *InnerI = dyn_cast<Instruction>(Inst.getOperand(0)); 256 auto *InnerOBO = cast<OverflowingBinaryOperator>(Inst.getOperand(0)); 257 bool SignedOverflow; 258 APInt NewC = C1->sadd_ov(*C2, SignedOverflow); 259 Inst.setOperand(0, X); 260 Inst.setOperand(1, ConstantInt::get(Inst.getType(), NewC)); 261 Inst.setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap() && 262 InnerOBO->hasNoUnsignedWrap()); 263 Inst.setHasNoSignedWrap(Inst.hasNoSignedWrap() && 264 InnerOBO->hasNoSignedWrap() && 265 !SignedOverflow); 266 if (InnerI && isInstructionTriviallyDead(InnerI)) 267 DeadInsts.emplace_back(InnerI); 268 } 269 } 270 } 271 // We can't do recursive deletion until we're done iterating, as we might 272 // have a phi which (potentially indirectly) uses instructions later in 273 // the block we're iterating through. 274 RecursivelyDeleteTriviallyDeadInstructions(DeadInsts); 275 } 276 } 277 278 /// Unroll the given loop by Count. The loop must be in LCSSA form. Unrolling 279 /// can only fail when the loop's latch block is not terminated by a conditional 280 /// branch instruction. However, if the trip count (and multiple) are not known, 281 /// loop unrolling will mostly produce more code that is no faster. 282 /// 283 /// If Runtime is true then UnrollLoop will try to insert a prologue or 284 /// epilogue that ensures the latch has a trip multiple of Count. UnrollLoop 285 /// will not runtime-unroll the loop if computing the run-time trip count will 286 /// be expensive and AllowExpensiveTripCount is false. 287 /// 288 /// The LoopInfo Analysis that is passed will be kept consistent. 289 /// 290 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and 291 /// DominatorTree if they are non-null. 292 /// 293 /// If RemainderLoop is non-null, it will receive the remainder loop (if 294 /// required and not fully unrolled). 295 LoopUnrollResult llvm::UnrollLoop(Loop *L, UnrollLoopOptions ULO, LoopInfo *LI, 296 ScalarEvolution *SE, DominatorTree *DT, 297 AssumptionCache *AC, 298 const TargetTransformInfo *TTI, 299 OptimizationRemarkEmitter *ORE, 300 bool PreserveLCSSA, Loop **RemainderLoop) { 301 assert(DT && "DomTree is required"); 302 303 if (!L->getLoopPreheader()) { 304 LLVM_DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n"); 305 return LoopUnrollResult::Unmodified; 306 } 307 308 if (!L->getLoopLatch()) { 309 LLVM_DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n"); 310 return LoopUnrollResult::Unmodified; 311 } 312 313 // Loops with indirectbr cannot be cloned. 314 if (!L->isSafeToClone()) { 315 LLVM_DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n"); 316 return LoopUnrollResult::Unmodified; 317 } 318 319 if (L->getHeader()->hasAddressTaken()) { 320 // The loop-rotate pass can be helpful to avoid this in many cases. 321 LLVM_DEBUG( 322 dbgs() << " Won't unroll loop: address of header block is taken.\n"); 323 return LoopUnrollResult::Unmodified; 324 } 325 326 assert(ULO.Count > 0); 327 328 // All these values should be taken only after peeling because they might have 329 // changed. 330 BasicBlock *Preheader = L->getLoopPreheader(); 331 BasicBlock *Header = L->getHeader(); 332 BasicBlock *LatchBlock = L->getLoopLatch(); 333 SmallVector<BasicBlock *, 4> ExitBlocks; 334 L->getExitBlocks(ExitBlocks); 335 std::vector<BasicBlock *> OriginalLoopBlocks = L->getBlocks(); 336 337 const unsigned MaxTripCount = SE->getSmallConstantMaxTripCount(L); 338 const bool MaxOrZero = SE->isBackedgeTakenCountMaxOrZero(L); 339 unsigned EstimatedLoopInvocationWeight = 0; 340 std::optional<unsigned> OriginalTripCount = 341 llvm::getLoopEstimatedTripCount(L, &EstimatedLoopInvocationWeight); 342 343 // Effectively "DCE" unrolled iterations that are beyond the max tripcount 344 // and will never be executed. 345 if (MaxTripCount && ULO.Count > MaxTripCount) 346 ULO.Count = MaxTripCount; 347 348 struct ExitInfo { 349 unsigned TripCount; 350 unsigned TripMultiple; 351 unsigned BreakoutTrip; 352 bool ExitOnTrue; 353 BasicBlock *FirstExitingBlock = nullptr; 354 SmallVector<BasicBlock *> ExitingBlocks; 355 }; 356 DenseMap<BasicBlock *, ExitInfo> ExitInfos; 357 SmallVector<BasicBlock *, 4> ExitingBlocks; 358 L->getExitingBlocks(ExitingBlocks); 359 for (auto *ExitingBlock : ExitingBlocks) { 360 // The folding code is not prepared to deal with non-branch instructions 361 // right now. 362 auto *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 363 if (!BI) 364 continue; 365 366 ExitInfo &Info = ExitInfos.try_emplace(ExitingBlock).first->second; 367 Info.TripCount = SE->getSmallConstantTripCount(L, ExitingBlock); 368 Info.TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock); 369 if (Info.TripCount != 0) { 370 Info.BreakoutTrip = Info.TripCount % ULO.Count; 371 Info.TripMultiple = 0; 372 } else { 373 Info.BreakoutTrip = Info.TripMultiple = 374 (unsigned)std::gcd(ULO.Count, Info.TripMultiple); 375 } 376 Info.ExitOnTrue = !L->contains(BI->getSuccessor(0)); 377 Info.ExitingBlocks.push_back(ExitingBlock); 378 LLVM_DEBUG(dbgs() << " Exiting block %" << ExitingBlock->getName() 379 << ": TripCount=" << Info.TripCount 380 << ", TripMultiple=" << Info.TripMultiple 381 << ", BreakoutTrip=" << Info.BreakoutTrip << "\n"); 382 } 383 384 // Are we eliminating the loop control altogether? Note that we can know 385 // we're eliminating the backedge without knowing exactly which iteration 386 // of the unrolled body exits. 387 const bool CompletelyUnroll = ULO.Count == MaxTripCount; 388 389 const bool PreserveOnlyFirst = CompletelyUnroll && MaxOrZero; 390 391 // There's no point in performing runtime unrolling if this unroll count 392 // results in a full unroll. 393 if (CompletelyUnroll) 394 ULO.Runtime = false; 395 396 // Go through all exits of L and see if there are any phi-nodes there. We just 397 // conservatively assume that they're inserted to preserve LCSSA form, which 398 // means that complete unrolling might break this form. We need to either fix 399 // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For 400 // now we just recompute LCSSA for the outer loop, but it should be possible 401 // to fix it in-place. 402 bool NeedToFixLCSSA = 403 PreserveLCSSA && CompletelyUnroll && 404 any_of(ExitBlocks, 405 [](const BasicBlock *BB) { return isa<PHINode>(BB->begin()); }); 406 407 // The current loop unroll pass can unroll loops that have 408 // (1) single latch; and 409 // (2a) latch is unconditional; or 410 // (2b) latch is conditional and is an exiting block 411 // FIXME: The implementation can be extended to work with more complicated 412 // cases, e.g. loops with multiple latches. 413 BranchInst *LatchBI = dyn_cast<BranchInst>(LatchBlock->getTerminator()); 414 415 // A conditional branch which exits the loop, which can be optimized to an 416 // unconditional branch in the unrolled loop in some cases. 417 bool LatchIsExiting = L->isLoopExiting(LatchBlock); 418 if (!LatchBI || (LatchBI->isConditional() && !LatchIsExiting)) { 419 LLVM_DEBUG( 420 dbgs() << "Can't unroll; a conditional latch must exit the loop"); 421 return LoopUnrollResult::Unmodified; 422 } 423 424 // Loops containing convergent instructions cannot use runtime unrolling, 425 // as the prologue/epilogue may add additional control-dependencies to 426 // convergent operations. 427 LLVM_DEBUG( 428 { 429 bool HasConvergent = false; 430 for (auto &BB : L->blocks()) 431 for (auto &I : *BB) 432 if (auto *CB = dyn_cast<CallBase>(&I)) 433 HasConvergent |= CB->isConvergent(); 434 assert((!HasConvergent || !ULO.Runtime) && 435 "Can't runtime unroll if loop contains a convergent operation."); 436 }); 437 438 bool EpilogProfitability = 439 UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog 440 : isEpilogProfitable(L); 441 442 if (ULO.Runtime && 443 !UnrollRuntimeLoopRemainder(L, ULO.Count, ULO.AllowExpensiveTripCount, 444 EpilogProfitability, ULO.UnrollRemainder, 445 ULO.ForgetAllSCEV, LI, SE, DT, AC, TTI, 446 PreserveLCSSA, RemainderLoop)) { 447 if (ULO.Force) 448 ULO.Runtime = false; 449 else { 450 LLVM_DEBUG(dbgs() << "Won't unroll; remainder loop could not be " 451 "generated when assuming runtime trip count\n"); 452 return LoopUnrollResult::Unmodified; 453 } 454 } 455 456 using namespace ore; 457 // Report the unrolling decision. 458 if (CompletelyUnroll) { 459 LLVM_DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName() 460 << " with trip count " << ULO.Count << "!\n"); 461 if (ORE) 462 ORE->emit([&]() { 463 return OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(), 464 L->getHeader()) 465 << "completely unrolled loop with " 466 << NV("UnrollCount", ULO.Count) << " iterations"; 467 }); 468 } else { 469 LLVM_DEBUG(dbgs() << "UNROLLING loop %" << Header->getName() << " by " 470 << ULO.Count); 471 if (ULO.Runtime) 472 LLVM_DEBUG(dbgs() << " with run-time trip count"); 473 LLVM_DEBUG(dbgs() << "!\n"); 474 475 if (ORE) 476 ORE->emit([&]() { 477 OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(), 478 L->getHeader()); 479 Diag << "unrolled loop by a factor of " << NV("UnrollCount", ULO.Count); 480 if (ULO.Runtime) 481 Diag << " with run-time trip count"; 482 return Diag; 483 }); 484 } 485 486 // We are going to make changes to this loop. SCEV may be keeping cached info 487 // about it, in particular about backedge taken count. The changes we make 488 // are guaranteed to invalidate this information for our loop. It is tempting 489 // to only invalidate the loop being unrolled, but it is incorrect as long as 490 // all exiting branches from all inner loops have impact on the outer loops, 491 // and if something changes inside them then any of outer loops may also 492 // change. When we forget outermost loop, we also forget all contained loops 493 // and this is what we need here. 494 if (SE) { 495 if (ULO.ForgetAllSCEV) 496 SE->forgetAllLoops(); 497 else { 498 SE->forgetTopmostLoop(L); 499 SE->forgetBlockAndLoopDispositions(); 500 } 501 } 502 503 if (!LatchIsExiting) 504 ++NumUnrolledNotLatch; 505 506 // For the first iteration of the loop, we should use the precloned values for 507 // PHI nodes. Insert associations now. 508 ValueToValueMapTy LastValueMap; 509 std::vector<PHINode*> OrigPHINode; 510 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 511 OrigPHINode.push_back(cast<PHINode>(I)); 512 } 513 514 std::vector<BasicBlock *> Headers; 515 std::vector<BasicBlock *> Latches; 516 Headers.push_back(Header); 517 Latches.push_back(LatchBlock); 518 519 // The current on-the-fly SSA update requires blocks to be processed in 520 // reverse postorder so that LastValueMap contains the correct value at each 521 // exit. 522 LoopBlocksDFS DFS(L); 523 DFS.perform(LI); 524 525 // Stash the DFS iterators before adding blocks to the loop. 526 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO(); 527 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO(); 528 529 std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks(); 530 531 // Loop Unrolling might create new loops. While we do preserve LoopInfo, we 532 // might break loop-simplified form for these loops (as they, e.g., would 533 // share the same exit blocks). We'll keep track of loops for which we can 534 // break this so that later we can re-simplify them. 535 SmallSetVector<Loop *, 4> LoopsToSimplify; 536 for (Loop *SubLoop : *L) 537 LoopsToSimplify.insert(SubLoop); 538 539 // When a FSDiscriminator is enabled, we don't need to add the multiply 540 // factors to the discriminators. 541 if (Header->getParent()->shouldEmitDebugInfoForProfiling() && 542 !EnableFSDiscriminator) 543 for (BasicBlock *BB : L->getBlocks()) 544 for (Instruction &I : *BB) 545 if (!I.isDebugOrPseudoInst()) 546 if (const DILocation *DIL = I.getDebugLoc()) { 547 auto NewDIL = DIL->cloneByMultiplyingDuplicationFactor(ULO.Count); 548 if (NewDIL) 549 I.setDebugLoc(*NewDIL); 550 else 551 LLVM_DEBUG(dbgs() 552 << "Failed to create new discriminator: " 553 << DIL->getFilename() << " Line: " << DIL->getLine()); 554 } 555 556 // Identify what noalias metadata is inside the loop: if it is inside the 557 // loop, the associated metadata must be cloned for each iteration. 558 SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes; 559 identifyNoAliasScopesToClone(L->getBlocks(), LoopLocalNoAliasDeclScopes); 560 561 // We place the unrolled iterations immediately after the original loop 562 // latch. This is a reasonable default placement if we don't have block 563 // frequencies, and if we do, well the layout will be adjusted later. 564 auto BlockInsertPt = std::next(LatchBlock->getIterator()); 565 for (unsigned It = 1; It != ULO.Count; ++It) { 566 SmallVector<BasicBlock *, 8> NewBlocks; 567 SmallDenseMap<const Loop *, Loop *, 4> NewLoops; 568 NewLoops[L] = L; 569 570 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { 571 ValueToValueMapTy VMap; 572 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It)); 573 Header->getParent()->insert(BlockInsertPt, New); 574 575 assert((*BB != Header || LI->getLoopFor(*BB) == L) && 576 "Header should not be in a sub-loop"); 577 // Tell LI about New. 578 const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops); 579 if (OldLoop) 580 LoopsToSimplify.insert(NewLoops[OldLoop]); 581 582 if (*BB == Header) 583 // Loop over all of the PHI nodes in the block, changing them to use 584 // the incoming values from the previous block. 585 for (PHINode *OrigPHI : OrigPHINode) { 586 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]); 587 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock); 588 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) 589 if (It > 1 && L->contains(InValI)) 590 InVal = LastValueMap[InValI]; 591 VMap[OrigPHI] = InVal; 592 NewPHI->eraseFromParent(); 593 } 594 595 // Update our running map of newest clones 596 LastValueMap[*BB] = New; 597 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end(); 598 VI != VE; ++VI) 599 LastValueMap[VI->first] = VI->second; 600 601 // Add phi entries for newly created values to all exit blocks. 602 for (BasicBlock *Succ : successors(*BB)) { 603 if (L->contains(Succ)) 604 continue; 605 for (PHINode &PHI : Succ->phis()) { 606 Value *Incoming = PHI.getIncomingValueForBlock(*BB); 607 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming); 608 if (It != LastValueMap.end()) 609 Incoming = It->second; 610 PHI.addIncoming(Incoming, New); 611 SE->forgetValue(&PHI); 612 } 613 } 614 // Keep track of new headers and latches as we create them, so that 615 // we can insert the proper branches later. 616 if (*BB == Header) 617 Headers.push_back(New); 618 if (*BB == LatchBlock) 619 Latches.push_back(New); 620 621 // Keep track of the exiting block and its successor block contained in 622 // the loop for the current iteration. 623 auto ExitInfoIt = ExitInfos.find(*BB); 624 if (ExitInfoIt != ExitInfos.end()) 625 ExitInfoIt->second.ExitingBlocks.push_back(New); 626 627 NewBlocks.push_back(New); 628 UnrolledLoopBlocks.push_back(New); 629 630 // Update DomTree: since we just copy the loop body, and each copy has a 631 // dedicated entry block (copy of the header block), this header's copy 632 // dominates all copied blocks. That means, dominance relations in the 633 // copied body are the same as in the original body. 634 if (*BB == Header) 635 DT->addNewBlock(New, Latches[It - 1]); 636 else { 637 auto BBDomNode = DT->getNode(*BB); 638 auto BBIDom = BBDomNode->getIDom(); 639 BasicBlock *OriginalBBIDom = BBIDom->getBlock(); 640 DT->addNewBlock( 641 New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)])); 642 } 643 } 644 645 // Remap all instructions in the most recent iteration 646 remapInstructionsInBlocks(NewBlocks, LastValueMap); 647 for (BasicBlock *NewBlock : NewBlocks) 648 for (Instruction &I : *NewBlock) 649 if (auto *II = dyn_cast<AssumeInst>(&I)) 650 AC->registerAssumption(II); 651 652 { 653 // Identify what other metadata depends on the cloned version. After 654 // cloning, replace the metadata with the corrected version for both 655 // memory instructions and noalias intrinsics. 656 std::string ext = (Twine("It") + Twine(It)).str(); 657 cloneAndAdaptNoAliasScopes(LoopLocalNoAliasDeclScopes, NewBlocks, 658 Header->getContext(), ext); 659 } 660 } 661 662 // Loop over the PHI nodes in the original block, setting incoming values. 663 for (PHINode *PN : OrigPHINode) { 664 if (CompletelyUnroll) { 665 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader)); 666 PN->eraseFromParent(); 667 } else if (ULO.Count > 1) { 668 Value *InVal = PN->removeIncomingValue(LatchBlock, false); 669 // If this value was defined in the loop, take the value defined by the 670 // last iteration of the loop. 671 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) { 672 if (L->contains(InValI)) 673 InVal = LastValueMap[InVal]; 674 } 675 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch"); 676 PN->addIncoming(InVal, Latches.back()); 677 } 678 } 679 680 // Connect latches of the unrolled iterations to the headers of the next 681 // iteration. Currently they point to the header of the same iteration. 682 for (unsigned i = 0, e = Latches.size(); i != e; ++i) { 683 unsigned j = (i + 1) % e; 684 Latches[i]->getTerminator()->replaceSuccessorWith(Headers[i], Headers[j]); 685 } 686 687 // Update dominators of blocks we might reach through exits. 688 // Immediate dominator of such block might change, because we add more 689 // routes which can lead to the exit: we can now reach it from the copied 690 // iterations too. 691 if (ULO.Count > 1) { 692 for (auto *BB : OriginalLoopBlocks) { 693 auto *BBDomNode = DT->getNode(BB); 694 SmallVector<BasicBlock *, 16> ChildrenToUpdate; 695 for (auto *ChildDomNode : BBDomNode->children()) { 696 auto *ChildBB = ChildDomNode->getBlock(); 697 if (!L->contains(ChildBB)) 698 ChildrenToUpdate.push_back(ChildBB); 699 } 700 // The new idom of the block will be the nearest common dominator 701 // of all copies of the previous idom. This is equivalent to the 702 // nearest common dominator of the previous idom and the first latch, 703 // which dominates all copies of the previous idom. 704 BasicBlock *NewIDom = DT->findNearestCommonDominator(BB, LatchBlock); 705 for (auto *ChildBB : ChildrenToUpdate) 706 DT->changeImmediateDominator(ChildBB, NewIDom); 707 } 708 } 709 710 assert(!UnrollVerifyDomtree || 711 DT->verify(DominatorTree::VerificationLevel::Fast)); 712 713 SmallVector<DominatorTree::UpdateType> DTUpdates; 714 auto SetDest = [&](BasicBlock *Src, bool WillExit, bool ExitOnTrue) { 715 auto *Term = cast<BranchInst>(Src->getTerminator()); 716 const unsigned Idx = ExitOnTrue ^ WillExit; 717 BasicBlock *Dest = Term->getSuccessor(Idx); 718 BasicBlock *DeadSucc = Term->getSuccessor(1-Idx); 719 720 // Remove predecessors from all non-Dest successors. 721 DeadSucc->removePredecessor(Src, /* KeepOneInputPHIs */ true); 722 723 // Replace the conditional branch with an unconditional one. 724 BranchInst::Create(Dest, Term); 725 Term->eraseFromParent(); 726 727 DTUpdates.emplace_back(DominatorTree::Delete, Src, DeadSucc); 728 }; 729 730 auto WillExit = [&](const ExitInfo &Info, unsigned i, unsigned j, 731 bool IsLatch) -> std::optional<bool> { 732 if (CompletelyUnroll) { 733 if (PreserveOnlyFirst) { 734 if (i == 0) 735 return std::nullopt; 736 return j == 0; 737 } 738 // Complete (but possibly inexact) unrolling 739 if (j == 0) 740 return true; 741 if (Info.TripCount && j != Info.TripCount) 742 return false; 743 return std::nullopt; 744 } 745 746 if (ULO.Runtime) { 747 // If runtime unrolling inserts a prologue, information about non-latch 748 // exits may be stale. 749 if (IsLatch && j != 0) 750 return false; 751 return std::nullopt; 752 } 753 754 if (j != Info.BreakoutTrip && 755 (Info.TripMultiple == 0 || j % Info.TripMultiple != 0)) { 756 // If we know the trip count or a multiple of it, we can safely use an 757 // unconditional branch for some iterations. 758 return false; 759 } 760 return std::nullopt; 761 }; 762 763 // Fold branches for iterations where we know that they will exit or not 764 // exit. 765 for (auto &Pair : ExitInfos) { 766 ExitInfo &Info = Pair.second; 767 for (unsigned i = 0, e = Info.ExitingBlocks.size(); i != e; ++i) { 768 // The branch destination. 769 unsigned j = (i + 1) % e; 770 bool IsLatch = Pair.first == LatchBlock; 771 std::optional<bool> KnownWillExit = WillExit(Info, i, j, IsLatch); 772 if (!KnownWillExit) { 773 if (!Info.FirstExitingBlock) 774 Info.FirstExitingBlock = Info.ExitingBlocks[i]; 775 continue; 776 } 777 778 // We don't fold known-exiting branches for non-latch exits here, 779 // because this ensures that both all loop blocks and all exit blocks 780 // remain reachable in the CFG. 781 // TODO: We could fold these branches, but it would require much more 782 // sophisticated updates to LoopInfo. 783 if (*KnownWillExit && !IsLatch) { 784 if (!Info.FirstExitingBlock) 785 Info.FirstExitingBlock = Info.ExitingBlocks[i]; 786 continue; 787 } 788 789 SetDest(Info.ExitingBlocks[i], *KnownWillExit, Info.ExitOnTrue); 790 } 791 } 792 793 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); 794 DomTreeUpdater *DTUToUse = &DTU; 795 if (ExitingBlocks.size() == 1 && ExitInfos.size() == 1) { 796 // Manually update the DT if there's a single exiting node. In that case 797 // there's a single exit node and it is sufficient to update the nodes 798 // immediately dominated by the original exiting block. They will become 799 // dominated by the first exiting block that leaves the loop after 800 // unrolling. Note that the CFG inside the loop does not change, so there's 801 // no need to update the DT inside the unrolled loop. 802 DTUToUse = nullptr; 803 auto &[OriginalExit, Info] = *ExitInfos.begin(); 804 if (!Info.FirstExitingBlock) 805 Info.FirstExitingBlock = Info.ExitingBlocks.back(); 806 for (auto *C : to_vector(DT->getNode(OriginalExit)->children())) { 807 if (L->contains(C->getBlock())) 808 continue; 809 C->setIDom(DT->getNode(Info.FirstExitingBlock)); 810 } 811 } else { 812 DTU.applyUpdates(DTUpdates); 813 } 814 815 // When completely unrolling, the last latch becomes unreachable. 816 if (!LatchIsExiting && CompletelyUnroll) { 817 // There is no need to update the DT here, because there must be a unique 818 // latch. Hence if the latch is not exiting it must directly branch back to 819 // the original loop header and does not dominate any nodes. 820 assert(LatchBlock->getSingleSuccessor() && "Loop with multiple latches?"); 821 changeToUnreachable(Latches.back()->getTerminator(), PreserveLCSSA); 822 } 823 824 // Merge adjacent basic blocks, if possible. 825 for (BasicBlock *Latch : Latches) { 826 BranchInst *Term = dyn_cast<BranchInst>(Latch->getTerminator()); 827 assert((Term || 828 (CompletelyUnroll && !LatchIsExiting && Latch == Latches.back())) && 829 "Need a branch as terminator, except when fully unrolling with " 830 "unconditional latch"); 831 if (Term && Term->isUnconditional()) { 832 BasicBlock *Dest = Term->getSuccessor(0); 833 BasicBlock *Fold = Dest->getUniquePredecessor(); 834 if (MergeBlockIntoPredecessor(Dest, /*DTU=*/DTUToUse, LI, 835 /*MSSAU=*/nullptr, /*MemDep=*/nullptr, 836 /*PredecessorWithTwoSuccessors=*/false, 837 DTUToUse ? nullptr : DT)) { 838 // Dest has been folded into Fold. Update our worklists accordingly. 839 std::replace(Latches.begin(), Latches.end(), Dest, Fold); 840 llvm::erase(UnrolledLoopBlocks, Dest); 841 } 842 } 843 } 844 845 if (DTUToUse) { 846 // Apply updates to the DomTree. 847 DT = &DTU.getDomTree(); 848 } 849 assert(!UnrollVerifyDomtree || 850 DT->verify(DominatorTree::VerificationLevel::Fast)); 851 852 // At this point, the code is well formed. We now simplify the unrolled loop, 853 // doing constant propagation and dead code elimination as we go. 854 simplifyLoopAfterUnroll(L, !CompletelyUnroll && ULO.Count > 1, LI, SE, DT, AC, 855 TTI); 856 857 NumCompletelyUnrolled += CompletelyUnroll; 858 ++NumUnrolled; 859 860 Loop *OuterL = L->getParentLoop(); 861 // Update LoopInfo if the loop is completely removed. 862 if (CompletelyUnroll) { 863 LI->erase(L); 864 // We shouldn't try to use `L` anymore. 865 L = nullptr; 866 } else if (OriginalTripCount) { 867 // Update the trip count. Note that the remainder has already logic 868 // computing it in `UnrollRuntimeLoopRemainder`. 869 setLoopEstimatedTripCount(L, *OriginalTripCount / ULO.Count, 870 EstimatedLoopInvocationWeight); 871 } 872 873 // LoopInfo should not be valid, confirm that. 874 if (UnrollVerifyLoopInfo) 875 LI->verify(*DT); 876 877 // After complete unrolling most of the blocks should be contained in OuterL. 878 // However, some of them might happen to be out of OuterL (e.g. if they 879 // precede a loop exit). In this case we might need to insert PHI nodes in 880 // order to preserve LCSSA form. 881 // We don't need to check this if we already know that we need to fix LCSSA 882 // form. 883 // TODO: For now we just recompute LCSSA for the outer loop in this case, but 884 // it should be possible to fix it in-place. 885 if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA) 886 NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI); 887 888 // Make sure that loop-simplify form is preserved. We want to simplify 889 // at least one layer outside of the loop that was unrolled so that any 890 // changes to the parent loop exposed by the unrolling are considered. 891 if (OuterL) { 892 // OuterL includes all loops for which we can break loop-simplify, so 893 // it's sufficient to simplify only it (it'll recursively simplify inner 894 // loops too). 895 if (NeedToFixLCSSA) { 896 // LCSSA must be performed on the outermost affected loop. The unrolled 897 // loop's last loop latch is guaranteed to be in the outermost loop 898 // after LoopInfo's been updated by LoopInfo::erase. 899 Loop *LatchLoop = LI->getLoopFor(Latches.back()); 900 Loop *FixLCSSALoop = OuterL; 901 if (!FixLCSSALoop->contains(LatchLoop)) 902 while (FixLCSSALoop->getParentLoop() != LatchLoop) 903 FixLCSSALoop = FixLCSSALoop->getParentLoop(); 904 905 formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE); 906 } else if (PreserveLCSSA) { 907 assert(OuterL->isLCSSAForm(*DT) && 908 "Loops should be in LCSSA form after loop-unroll."); 909 } 910 911 // TODO: That potentially might be compile-time expensive. We should try 912 // to fix the loop-simplified form incrementally. 913 simplifyLoop(OuterL, DT, LI, SE, AC, nullptr, PreserveLCSSA); 914 } else { 915 // Simplify loops for which we might've broken loop-simplify form. 916 for (Loop *SubLoop : LoopsToSimplify) 917 simplifyLoop(SubLoop, DT, LI, SE, AC, nullptr, PreserveLCSSA); 918 } 919 920 return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled 921 : LoopUnrollResult::PartiallyUnrolled; 922 } 923 924 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata 925 /// node with the given name (for example, "llvm.loop.unroll.count"). If no 926 /// such metadata node exists, then nullptr is returned. 927 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) { 928 // First operand should refer to the loop id itself. 929 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 930 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 931 932 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 933 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 934 if (!MD) 935 continue; 936 937 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 938 if (!S) 939 continue; 940 941 if (Name.equals(S->getString())) 942 return MD; 943 } 944 return nullptr; 945 } 946