1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements some loop unrolling utilities. It does not define any 11 // actual pass or policy, but provides a single function to perform loop 12 // unrolling. 13 // 14 // The process of unrolling can produce extraneous basic blocks linked with 15 // unconditional branches. This will be corrected in the future. 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/Analysis/AssumptionCache.h" 22 #include "llvm/Analysis/InstructionSimplify.h" 23 #include "llvm/Analysis/LoopIterator.h" 24 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 25 #include "llvm/Analysis/ScalarEvolution.h" 26 #include "llvm/Transforms/Utils/Local.h" 27 #include "llvm/IR/BasicBlock.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/DebugInfoMetadata.h" 30 #include "llvm/IR/Dominators.h" 31 #include "llvm/IR/IntrinsicInst.h" 32 #include "llvm/IR/LLVMContext.h" 33 #include "llvm/Support/Debug.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 36 #include "llvm/Transforms/Utils/Cloning.h" 37 #include "llvm/Transforms/Utils/LoopSimplify.h" 38 #include "llvm/Transforms/Utils/LoopUtils.h" 39 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 40 #include "llvm/Transforms/Utils/UnrollLoop.h" 41 using namespace llvm; 42 43 #define DEBUG_TYPE "loop-unroll" 44 45 // TODO: Should these be here or in LoopUnroll? 46 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled"); 47 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)"); 48 49 static cl::opt<bool> 50 UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden, 51 cl::desc("Allow runtime unrolled loops to be unrolled " 52 "with epilog instead of prolog.")); 53 54 static cl::opt<bool> 55 UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden, 56 cl::desc("Verify domtree after unrolling"), 57 #ifdef NDEBUG 58 cl::init(false) 59 #else 60 cl::init(true) 61 #endif 62 ); 63 64 /// Convert the instruction operands from referencing the current values into 65 /// those specified by VMap. 66 void llvm::remapInstruction(Instruction *I, ValueToValueMapTy &VMap) { 67 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 68 Value *Op = I->getOperand(op); 69 70 // Unwrap arguments of dbg.value intrinsics. 71 bool Wrapped = false; 72 if (auto *V = dyn_cast<MetadataAsValue>(Op)) 73 if (auto *Unwrapped = dyn_cast<ValueAsMetadata>(V->getMetadata())) { 74 Op = Unwrapped->getValue(); 75 Wrapped = true; 76 } 77 78 auto wrap = [&](Value *V) { 79 auto &C = I->getContext(); 80 return Wrapped ? MetadataAsValue::get(C, ValueAsMetadata::get(V)) : V; 81 }; 82 83 ValueToValueMapTy::iterator It = VMap.find(Op); 84 if (It != VMap.end()) 85 I->setOperand(op, wrap(It->second)); 86 } 87 88 if (PHINode *PN = dyn_cast<PHINode>(I)) { 89 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 90 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i)); 91 if (It != VMap.end()) 92 PN->setIncomingBlock(i, cast<BasicBlock>(It->second)); 93 } 94 } 95 } 96 97 /// Folds a basic block into its predecessor if it only has one predecessor, and 98 /// that predecessor only has one successor. 99 /// The LoopInfo Analysis that is passed will be kept consistent. 100 BasicBlock *llvm::foldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI, 101 ScalarEvolution *SE, 102 DominatorTree *DT) { 103 // Merge basic blocks into their predecessor if there is only one distinct 104 // pred, and if there is only one distinct successor of the predecessor, and 105 // if there are no PHI nodes. 106 BasicBlock *OnlyPred = BB->getSinglePredecessor(); 107 if (!OnlyPred) return nullptr; 108 109 if (OnlyPred->getTerminator()->getNumSuccessors() != 1) 110 return nullptr; 111 112 LLVM_DEBUG(dbgs() << "Merging: " << BB->getName() << " into " 113 << OnlyPred->getName() << "\n"); 114 115 // Resolve any PHI nodes at the start of the block. They are all 116 // guaranteed to have exactly one entry if they exist, unless there are 117 // multiple duplicate (but guaranteed to be equal) entries for the 118 // incoming edges. This occurs when there are multiple edges from 119 // OnlyPred to OnlySucc. 120 FoldSingleEntryPHINodes(BB); 121 122 // Delete the unconditional branch from the predecessor... 123 OnlyPred->getInstList().pop_back(); 124 125 // Make all PHI nodes that referred to BB now refer to Pred as their 126 // source... 127 BB->replaceAllUsesWith(OnlyPred); 128 129 // Move all definitions in the successor to the predecessor... 130 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList()); 131 132 // OldName will be valid until erased. 133 StringRef OldName = BB->getName(); 134 135 // Erase the old block and update dominator info. 136 if (DT) 137 if (DomTreeNode *DTN = DT->getNode(BB)) { 138 DomTreeNode *PredDTN = DT->getNode(OnlyPred); 139 SmallVector<DomTreeNode *, 8> Children(DTN->begin(), DTN->end()); 140 for (auto *DI : Children) 141 DT->changeImmediateDominator(DI, PredDTN); 142 143 DT->eraseNode(BB); 144 } 145 146 LI->removeBlock(BB); 147 148 // Inherit predecessor's name if it exists... 149 if (!OldName.empty() && !OnlyPred->hasName()) 150 OnlyPred->setName(OldName); 151 152 BB->eraseFromParent(); 153 154 return OnlyPred; 155 } 156 157 /// Check if unrolling created a situation where we need to insert phi nodes to 158 /// preserve LCSSA form. 159 /// \param Blocks is a vector of basic blocks representing unrolled loop. 160 /// \param L is the outer loop. 161 /// It's possible that some of the blocks are in L, and some are not. In this 162 /// case, if there is a use is outside L, and definition is inside L, we need to 163 /// insert a phi-node, otherwise LCSSA will be broken. 164 /// The function is just a helper function for llvm::UnrollLoop that returns 165 /// true if this situation occurs, indicating that LCSSA needs to be fixed. 166 static bool needToInsertPhisForLCSSA(Loop *L, std::vector<BasicBlock *> Blocks, 167 LoopInfo *LI) { 168 for (BasicBlock *BB : Blocks) { 169 if (LI->getLoopFor(BB) == L) 170 continue; 171 for (Instruction &I : *BB) { 172 for (Use &U : I.operands()) { 173 if (auto Def = dyn_cast<Instruction>(U)) { 174 Loop *DefLoop = LI->getLoopFor(Def->getParent()); 175 if (!DefLoop) 176 continue; 177 if (DefLoop->contains(L)) 178 return true; 179 } 180 } 181 } 182 } 183 return false; 184 } 185 186 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary 187 /// and adds a mapping from the original loop to the new loop to NewLoops. 188 /// Returns nullptr if no new loop was created and a pointer to the 189 /// original loop OriginalBB was part of otherwise. 190 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB, 191 BasicBlock *ClonedBB, LoopInfo *LI, 192 NewLoopsMap &NewLoops) { 193 // Figure out which loop New is in. 194 const Loop *OldLoop = LI->getLoopFor(OriginalBB); 195 assert(OldLoop && "Should (at least) be in the loop being unrolled!"); 196 197 Loop *&NewLoop = NewLoops[OldLoop]; 198 if (!NewLoop) { 199 // Found a new sub-loop. 200 assert(OriginalBB == OldLoop->getHeader() && 201 "Header should be first in RPO"); 202 203 NewLoop = LI->AllocateLoop(); 204 Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop()); 205 206 if (NewLoopParent) 207 NewLoopParent->addChildLoop(NewLoop); 208 else 209 LI->addTopLevelLoop(NewLoop); 210 211 NewLoop->addBasicBlockToLoop(ClonedBB, *LI); 212 return OldLoop; 213 } else { 214 NewLoop->addBasicBlockToLoop(ClonedBB, *LI); 215 return nullptr; 216 } 217 } 218 219 /// The function chooses which type of unroll (epilog or prolog) is more 220 /// profitabale. 221 /// Epilog unroll is more profitable when there is PHI that starts from 222 /// constant. In this case epilog will leave PHI start from constant, 223 /// but prolog will convert it to non-constant. 224 /// 225 /// loop: 226 /// PN = PHI [I, Latch], [CI, PreHeader] 227 /// I = foo(PN) 228 /// ... 229 /// 230 /// Epilog unroll case. 231 /// loop: 232 /// PN = PHI [I2, Latch], [CI, PreHeader] 233 /// I1 = foo(PN) 234 /// I2 = foo(I1) 235 /// ... 236 /// Prolog unroll case. 237 /// NewPN = PHI [PrologI, Prolog], [CI, PreHeader] 238 /// loop: 239 /// PN = PHI [I2, Latch], [NewPN, PreHeader] 240 /// I1 = foo(PN) 241 /// I2 = foo(I1) 242 /// ... 243 /// 244 static bool isEpilogProfitable(Loop *L) { 245 BasicBlock *PreHeader = L->getLoopPreheader(); 246 BasicBlock *Header = L->getHeader(); 247 assert(PreHeader && Header); 248 for (const PHINode &PN : Header->phis()) { 249 if (isa<ConstantInt>(PN.getIncomingValueForBlock(PreHeader))) 250 return true; 251 } 252 return false; 253 } 254 255 /// Perform some cleanup and simplifications on loops after unrolling. It is 256 /// useful to simplify the IV's in the new loop, as well as do a quick 257 /// simplify/dce pass of the instructions. 258 void llvm::simplifyLoopAfterUnroll(Loop *L, bool SimplifyIVs, LoopInfo *LI, 259 ScalarEvolution *SE, DominatorTree *DT, 260 AssumptionCache *AC) { 261 // Simplify any new induction variables in the partially unrolled loop. 262 if (SE && SimplifyIVs) { 263 SmallVector<WeakTrackingVH, 16> DeadInsts; 264 simplifyLoopIVs(L, SE, DT, LI, DeadInsts); 265 266 // Aggressively clean up dead instructions that simplifyLoopIVs already 267 // identified. Any remaining should be cleaned up below. 268 while (!DeadInsts.empty()) 269 if (Instruction *Inst = 270 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) 271 RecursivelyDeleteTriviallyDeadInstructions(Inst); 272 } 273 274 // At this point, the code is well formed. We now do a quick sweep over the 275 // inserted code, doing constant propagation and dead code elimination as we 276 // go. 277 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 278 const std::vector<BasicBlock *> &NewLoopBlocks = L->getBlocks(); 279 for (BasicBlock *BB : NewLoopBlocks) { 280 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { 281 Instruction *Inst = &*I++; 282 283 if (Value *V = SimplifyInstruction(Inst, {DL, nullptr, DT, AC})) 284 if (LI->replacementPreservesLCSSAForm(Inst, V)) 285 Inst->replaceAllUsesWith(V); 286 if (isInstructionTriviallyDead(Inst)) 287 BB->getInstList().erase(Inst); 288 } 289 } 290 291 // TODO: after peeling or unrolling, previously loop variant conditions are 292 // likely to fold to constants, eagerly propagating those here will require 293 // fewer cleanup passes to be run. Alternatively, a LoopEarlyCSE might be 294 // appropriate. 295 } 296 297 /// Unroll the given loop by Count. The loop must be in LCSSA form. Unrolling 298 /// can only fail when the loop's latch block is not terminated by a conditional 299 /// branch instruction. However, if the trip count (and multiple) are not known, 300 /// loop unrolling will mostly produce more code that is no faster. 301 /// 302 /// TripCount is the upper bound of the iteration on which control exits 303 /// LatchBlock. Control may exit the loop prior to TripCount iterations either 304 /// via an early branch in other loop block or via LatchBlock terminator. This 305 /// is relaxed from the general definition of trip count which is the number of 306 /// times the loop header executes. Note that UnrollLoop assumes that the loop 307 /// counter test is in LatchBlock in order to remove unnecesssary instances of 308 /// the test. If control can exit the loop from the LatchBlock's terminator 309 /// prior to TripCount iterations, flag PreserveCondBr needs to be set. 310 /// 311 /// PreserveCondBr indicates whether the conditional branch of the LatchBlock 312 /// needs to be preserved. It is needed when we use trip count upper bound to 313 /// fully unroll the loop. If PreserveOnlyFirst is also set then only the first 314 /// conditional branch needs to be preserved. 315 /// 316 /// Similarly, TripMultiple divides the number of times that the LatchBlock may 317 /// execute without exiting the loop. 318 /// 319 /// If AllowRuntime is true then UnrollLoop will consider unrolling loops that 320 /// have a runtime (i.e. not compile time constant) trip count. Unrolling these 321 /// loops require a unroll "prologue" that runs "RuntimeTripCount % Count" 322 /// iterations before branching into the unrolled loop. UnrollLoop will not 323 /// runtime-unroll the loop if computing RuntimeTripCount will be expensive and 324 /// AllowExpensiveTripCount is false. 325 /// 326 /// If we want to perform PGO-based loop peeling, PeelCount is set to the 327 /// number of iterations we want to peel off. 328 /// 329 /// The LoopInfo Analysis that is passed will be kept consistent. 330 /// 331 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and 332 /// DominatorTree if they are non-null. 333 LoopUnrollResult llvm::UnrollLoop( 334 Loop *L, unsigned Count, unsigned TripCount, bool Force, bool AllowRuntime, 335 bool AllowExpensiveTripCount, bool PreserveCondBr, bool PreserveOnlyFirst, 336 unsigned TripMultiple, unsigned PeelCount, bool UnrollRemainder, 337 LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC, 338 OptimizationRemarkEmitter *ORE, bool PreserveLCSSA) { 339 340 BasicBlock *Preheader = L->getLoopPreheader(); 341 if (!Preheader) { 342 LLVM_DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n"); 343 return LoopUnrollResult::Unmodified; 344 } 345 346 BasicBlock *LatchBlock = L->getLoopLatch(); 347 if (!LatchBlock) { 348 LLVM_DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n"); 349 return LoopUnrollResult::Unmodified; 350 } 351 352 // Loops with indirectbr cannot be cloned. 353 if (!L->isSafeToClone()) { 354 LLVM_DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n"); 355 return LoopUnrollResult::Unmodified; 356 } 357 358 // The current loop unroll pass can only unroll loops with a single latch 359 // that's a conditional branch exiting the loop. 360 // FIXME: The implementation can be extended to work with more complicated 361 // cases, e.g. loops with multiple latches. 362 BasicBlock *Header = L->getHeader(); 363 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator()); 364 365 if (!BI || BI->isUnconditional()) { 366 // The loop-rotate pass can be helpful to avoid this in many cases. 367 LLVM_DEBUG( 368 dbgs() 369 << " Can't unroll; loop not terminated by a conditional branch.\n"); 370 return LoopUnrollResult::Unmodified; 371 } 372 373 auto CheckSuccessors = [&](unsigned S1, unsigned S2) { 374 return BI->getSuccessor(S1) == Header && !L->contains(BI->getSuccessor(S2)); 375 }; 376 377 if (!CheckSuccessors(0, 1) && !CheckSuccessors(1, 0)) { 378 LLVM_DEBUG(dbgs() << "Can't unroll; only loops with one conditional latch" 379 " exiting the loop can be unrolled\n"); 380 return LoopUnrollResult::Unmodified; 381 } 382 383 if (Header->hasAddressTaken()) { 384 // The loop-rotate pass can be helpful to avoid this in many cases. 385 LLVM_DEBUG( 386 dbgs() << " Won't unroll loop: address of header block is taken.\n"); 387 return LoopUnrollResult::Unmodified; 388 } 389 390 if (TripCount != 0) 391 LLVM_DEBUG(dbgs() << " Trip Count = " << TripCount << "\n"); 392 if (TripMultiple != 1) 393 LLVM_DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n"); 394 395 // Effectively "DCE" unrolled iterations that are beyond the tripcount 396 // and will never be executed. 397 if (TripCount != 0 && Count > TripCount) 398 Count = TripCount; 399 400 // Don't enter the unroll code if there is nothing to do. 401 if (TripCount == 0 && Count < 2 && PeelCount == 0) { 402 LLVM_DEBUG(dbgs() << "Won't unroll; almost nothing to do\n"); 403 return LoopUnrollResult::Unmodified; 404 } 405 406 assert(Count > 0); 407 assert(TripMultiple > 0); 408 assert(TripCount == 0 || TripCount % TripMultiple == 0); 409 410 // Are we eliminating the loop control altogether? 411 bool CompletelyUnroll = Count == TripCount; 412 SmallVector<BasicBlock *, 4> ExitBlocks; 413 L->getExitBlocks(ExitBlocks); 414 std::vector<BasicBlock*> OriginalLoopBlocks = L->getBlocks(); 415 416 // Go through all exits of L and see if there are any phi-nodes there. We just 417 // conservatively assume that they're inserted to preserve LCSSA form, which 418 // means that complete unrolling might break this form. We need to either fix 419 // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For 420 // now we just recompute LCSSA for the outer loop, but it should be possible 421 // to fix it in-place. 422 bool NeedToFixLCSSA = PreserveLCSSA && CompletelyUnroll && 423 any_of(ExitBlocks, [](const BasicBlock *BB) { 424 return isa<PHINode>(BB->begin()); 425 }); 426 427 // We assume a run-time trip count if the compiler cannot 428 // figure out the loop trip count and the unroll-runtime 429 // flag is specified. 430 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime); 431 432 assert((!RuntimeTripCount || !PeelCount) && 433 "Did not expect runtime trip-count unrolling " 434 "and peeling for the same loop"); 435 436 bool Peeled = false; 437 if (PeelCount) { 438 Peeled = peelLoop(L, PeelCount, LI, SE, DT, AC, PreserveLCSSA); 439 440 // Successful peeling may result in a change in the loop preheader/trip 441 // counts. If we later unroll the loop, we want these to be updated. 442 if (Peeled) { 443 BasicBlock *ExitingBlock = L->getExitingBlock(); 444 assert(ExitingBlock && "Loop without exiting block?"); 445 Preheader = L->getLoopPreheader(); 446 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock); 447 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock); 448 } 449 } 450 451 // Loops containing convergent instructions must have a count that divides 452 // their TripMultiple. 453 LLVM_DEBUG( 454 { 455 bool HasConvergent = false; 456 for (auto &BB : L->blocks()) 457 for (auto &I : *BB) 458 if (auto CS = CallSite(&I)) 459 HasConvergent |= CS.isConvergent(); 460 assert((!HasConvergent || TripMultiple % Count == 0) && 461 "Unroll count must divide trip multiple if loop contains a " 462 "convergent operation."); 463 }); 464 465 bool EpilogProfitability = 466 UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog 467 : isEpilogProfitable(L); 468 469 if (RuntimeTripCount && TripMultiple % Count != 0 && 470 !UnrollRuntimeLoopRemainder(L, Count, AllowExpensiveTripCount, 471 EpilogProfitability, UnrollRemainder, LI, SE, 472 DT, AC, PreserveLCSSA)) { 473 if (Force) 474 RuntimeTripCount = false; 475 else { 476 LLVM_DEBUG(dbgs() << "Won't unroll; remainder loop could not be " 477 "generated when assuming runtime trip count\n"); 478 return LoopUnrollResult::Unmodified; 479 } 480 } 481 482 // If we know the trip count, we know the multiple... 483 unsigned BreakoutTrip = 0; 484 if (TripCount != 0) { 485 BreakoutTrip = TripCount % Count; 486 TripMultiple = 0; 487 } else { 488 // Figure out what multiple to use. 489 BreakoutTrip = TripMultiple = 490 (unsigned)GreatestCommonDivisor64(Count, TripMultiple); 491 } 492 493 using namespace ore; 494 // Report the unrolling decision. 495 if (CompletelyUnroll) { 496 LLVM_DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName() 497 << " with trip count " << TripCount << "!\n"); 498 if (ORE) 499 ORE->emit([&]() { 500 return OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(), 501 L->getHeader()) 502 << "completely unrolled loop with " 503 << NV("UnrollCount", TripCount) << " iterations"; 504 }); 505 } else if (PeelCount) { 506 LLVM_DEBUG(dbgs() << "PEELING loop %" << Header->getName() 507 << " with iteration count " << PeelCount << "!\n"); 508 if (ORE) 509 ORE->emit([&]() { 510 return OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(), 511 L->getHeader()) 512 << " peeled loop by " << NV("PeelCount", PeelCount) 513 << " iterations"; 514 }); 515 } else { 516 auto DiagBuilder = [&]() { 517 OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(), 518 L->getHeader()); 519 return Diag << "unrolled loop by a factor of " 520 << NV("UnrollCount", Count); 521 }; 522 523 LLVM_DEBUG(dbgs() << "UNROLLING loop %" << Header->getName() << " by " 524 << Count); 525 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) { 526 LLVM_DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip); 527 if (ORE) 528 ORE->emit([&]() { 529 return DiagBuilder() << " with a breakout at trip " 530 << NV("BreakoutTrip", BreakoutTrip); 531 }); 532 } else if (TripMultiple != 1) { 533 LLVM_DEBUG(dbgs() << " with " << TripMultiple << " trips per branch"); 534 if (ORE) 535 ORE->emit([&]() { 536 return DiagBuilder() << " with " << NV("TripMultiple", TripMultiple) 537 << " trips per branch"; 538 }); 539 } else if (RuntimeTripCount) { 540 LLVM_DEBUG(dbgs() << " with run-time trip count"); 541 if (ORE) 542 ORE->emit( 543 [&]() { return DiagBuilder() << " with run-time trip count"; }); 544 } 545 LLVM_DEBUG(dbgs() << "!\n"); 546 } 547 548 // We are going to make changes to this loop. SCEV may be keeping cached info 549 // about it, in particular about backedge taken count. The changes we make 550 // are guaranteed to invalidate this information for our loop. It is tempting 551 // to only invalidate the loop being unrolled, but it is incorrect as long as 552 // all exiting branches from all inner loops have impact on the outer loops, 553 // and if something changes inside them then any of outer loops may also 554 // change. When we forget outermost loop, we also forget all contained loops 555 // and this is what we need here. 556 if (SE) 557 SE->forgetTopmostLoop(L); 558 559 bool ContinueOnTrue = L->contains(BI->getSuccessor(0)); 560 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue); 561 562 // For the first iteration of the loop, we should use the precloned values for 563 // PHI nodes. Insert associations now. 564 ValueToValueMapTy LastValueMap; 565 std::vector<PHINode*> OrigPHINode; 566 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 567 OrigPHINode.push_back(cast<PHINode>(I)); 568 } 569 570 std::vector<BasicBlock*> Headers; 571 std::vector<BasicBlock*> Latches; 572 Headers.push_back(Header); 573 Latches.push_back(LatchBlock); 574 575 // The current on-the-fly SSA update requires blocks to be processed in 576 // reverse postorder so that LastValueMap contains the correct value at each 577 // exit. 578 LoopBlocksDFS DFS(L); 579 DFS.perform(LI); 580 581 // Stash the DFS iterators before adding blocks to the loop. 582 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO(); 583 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO(); 584 585 std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks(); 586 587 // Loop Unrolling might create new loops. While we do preserve LoopInfo, we 588 // might break loop-simplified form for these loops (as they, e.g., would 589 // share the same exit blocks). We'll keep track of loops for which we can 590 // break this so that later we can re-simplify them. 591 SmallSetVector<Loop *, 4> LoopsToSimplify; 592 for (Loop *SubLoop : *L) 593 LoopsToSimplify.insert(SubLoop); 594 595 if (Header->getParent()->isDebugInfoForProfiling()) 596 for (BasicBlock *BB : L->getBlocks()) 597 for (Instruction &I : *BB) 598 if (!isa<DbgInfoIntrinsic>(&I)) 599 if (const DILocation *DIL = I.getDebugLoc()) 600 I.setDebugLoc(DIL->cloneWithDuplicationFactor(Count)); 601 602 for (unsigned It = 1; It != Count; ++It) { 603 std::vector<BasicBlock*> NewBlocks; 604 SmallDenseMap<const Loop *, Loop *, 4> NewLoops; 605 NewLoops[L] = L; 606 607 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { 608 ValueToValueMapTy VMap; 609 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It)); 610 Header->getParent()->getBasicBlockList().push_back(New); 611 612 assert((*BB != Header || LI->getLoopFor(*BB) == L) && 613 "Header should not be in a sub-loop"); 614 // Tell LI about New. 615 const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops); 616 if (OldLoop) 617 LoopsToSimplify.insert(NewLoops[OldLoop]); 618 619 if (*BB == Header) 620 // Loop over all of the PHI nodes in the block, changing them to use 621 // the incoming values from the previous block. 622 for (PHINode *OrigPHI : OrigPHINode) { 623 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]); 624 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock); 625 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) 626 if (It > 1 && L->contains(InValI)) 627 InVal = LastValueMap[InValI]; 628 VMap[OrigPHI] = InVal; 629 New->getInstList().erase(NewPHI); 630 } 631 632 // Update our running map of newest clones 633 LastValueMap[*BB] = New; 634 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end(); 635 VI != VE; ++VI) 636 LastValueMap[VI->first] = VI->second; 637 638 // Add phi entries for newly created values to all exit blocks. 639 for (BasicBlock *Succ : successors(*BB)) { 640 if (L->contains(Succ)) 641 continue; 642 for (PHINode &PHI : Succ->phis()) { 643 Value *Incoming = PHI.getIncomingValueForBlock(*BB); 644 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming); 645 if (It != LastValueMap.end()) 646 Incoming = It->second; 647 PHI.addIncoming(Incoming, New); 648 } 649 } 650 // Keep track of new headers and latches as we create them, so that 651 // we can insert the proper branches later. 652 if (*BB == Header) 653 Headers.push_back(New); 654 if (*BB == LatchBlock) 655 Latches.push_back(New); 656 657 NewBlocks.push_back(New); 658 UnrolledLoopBlocks.push_back(New); 659 660 // Update DomTree: since we just copy the loop body, and each copy has a 661 // dedicated entry block (copy of the header block), this header's copy 662 // dominates all copied blocks. That means, dominance relations in the 663 // copied body are the same as in the original body. 664 if (DT) { 665 if (*BB == Header) 666 DT->addNewBlock(New, Latches[It - 1]); 667 else { 668 auto BBDomNode = DT->getNode(*BB); 669 auto BBIDom = BBDomNode->getIDom(); 670 BasicBlock *OriginalBBIDom = BBIDom->getBlock(); 671 DT->addNewBlock( 672 New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)])); 673 } 674 } 675 } 676 677 // Remap all instructions in the most recent iteration 678 for (BasicBlock *NewBlock : NewBlocks) { 679 for (Instruction &I : *NewBlock) { 680 ::remapInstruction(&I, LastValueMap); 681 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 682 if (II->getIntrinsicID() == Intrinsic::assume) 683 AC->registerAssumption(II); 684 } 685 } 686 } 687 688 // Loop over the PHI nodes in the original block, setting incoming values. 689 for (PHINode *PN : OrigPHINode) { 690 if (CompletelyUnroll) { 691 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader)); 692 Header->getInstList().erase(PN); 693 } 694 else if (Count > 1) { 695 Value *InVal = PN->removeIncomingValue(LatchBlock, false); 696 // If this value was defined in the loop, take the value defined by the 697 // last iteration of the loop. 698 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) { 699 if (L->contains(InValI)) 700 InVal = LastValueMap[InVal]; 701 } 702 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch"); 703 PN->addIncoming(InVal, Latches.back()); 704 } 705 } 706 707 // Now that all the basic blocks for the unrolled iterations are in place, 708 // set up the branches to connect them. 709 for (unsigned i = 0, e = Latches.size(); i != e; ++i) { 710 // The original branch was replicated in each unrolled iteration. 711 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator()); 712 713 // The branch destination. 714 unsigned j = (i + 1) % e; 715 BasicBlock *Dest = Headers[j]; 716 bool NeedConditional = true; 717 718 if (RuntimeTripCount && j != 0) { 719 NeedConditional = false; 720 } 721 722 // For a complete unroll, make the last iteration end with a branch 723 // to the exit block. 724 if (CompletelyUnroll) { 725 if (j == 0) 726 Dest = LoopExit; 727 // If using trip count upper bound to completely unroll, we need to keep 728 // the conditional branch except the last one because the loop may exit 729 // after any iteration. 730 assert(NeedConditional && 731 "NeedCondition cannot be modified by both complete " 732 "unrolling and runtime unrolling"); 733 NeedConditional = (PreserveCondBr && j && !(PreserveOnlyFirst && i != 0)); 734 } else if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) { 735 // If we know the trip count or a multiple of it, we can safely use an 736 // unconditional branch for some iterations. 737 NeedConditional = false; 738 } 739 740 if (NeedConditional) { 741 // Update the conditional branch's successor for the following 742 // iteration. 743 Term->setSuccessor(!ContinueOnTrue, Dest); 744 } else { 745 // Remove phi operands at this loop exit 746 if (Dest != LoopExit) { 747 BasicBlock *BB = Latches[i]; 748 for (BasicBlock *Succ: successors(BB)) { 749 if (Succ == Headers[i]) 750 continue; 751 for (PHINode &Phi : Succ->phis()) 752 Phi.removeIncomingValue(BB, false); 753 } 754 } 755 // Replace the conditional branch with an unconditional one. 756 BranchInst::Create(Dest, Term); 757 Term->eraseFromParent(); 758 } 759 } 760 761 // Update dominators of blocks we might reach through exits. 762 // Immediate dominator of such block might change, because we add more 763 // routes which can lead to the exit: we can now reach it from the copied 764 // iterations too. 765 if (DT && Count > 1) { 766 for (auto *BB : OriginalLoopBlocks) { 767 auto *BBDomNode = DT->getNode(BB); 768 SmallVector<BasicBlock *, 16> ChildrenToUpdate; 769 for (auto *ChildDomNode : BBDomNode->getChildren()) { 770 auto *ChildBB = ChildDomNode->getBlock(); 771 if (!L->contains(ChildBB)) 772 ChildrenToUpdate.push_back(ChildBB); 773 } 774 BasicBlock *NewIDom; 775 if (BB == LatchBlock) { 776 // The latch is special because we emit unconditional branches in 777 // some cases where the original loop contained a conditional branch. 778 // Since the latch is always at the bottom of the loop, if the latch 779 // dominated an exit before unrolling, the new dominator of that exit 780 // must also be a latch. Specifically, the dominator is the first 781 // latch which ends in a conditional branch, or the last latch if 782 // there is no such latch. 783 NewIDom = Latches.back(); 784 for (BasicBlock *IterLatch : Latches) { 785 TerminatorInst *Term = IterLatch->getTerminator(); 786 if (isa<BranchInst>(Term) && cast<BranchInst>(Term)->isConditional()) { 787 NewIDom = IterLatch; 788 break; 789 } 790 } 791 } else { 792 // The new idom of the block will be the nearest common dominator 793 // of all copies of the previous idom. This is equivalent to the 794 // nearest common dominator of the previous idom and the first latch, 795 // which dominates all copies of the previous idom. 796 NewIDom = DT->findNearestCommonDominator(BB, LatchBlock); 797 } 798 for (auto *ChildBB : ChildrenToUpdate) 799 DT->changeImmediateDominator(ChildBB, NewIDom); 800 } 801 } 802 803 assert(!DT || !UnrollVerifyDomtree || 804 DT->verify(DominatorTree::VerificationLevel::Fast)); 805 806 // Merge adjacent basic blocks, if possible. 807 for (BasicBlock *Latch : Latches) { 808 BranchInst *Term = cast<BranchInst>(Latch->getTerminator()); 809 if (Term->isUnconditional()) { 810 BasicBlock *Dest = Term->getSuccessor(0); 811 if (BasicBlock *Fold = foldBlockIntoPredecessor(Dest, LI, SE, DT)) { 812 // Dest has been folded into Fold. Update our worklists accordingly. 813 std::replace(Latches.begin(), Latches.end(), Dest, Fold); 814 UnrolledLoopBlocks.erase(std::remove(UnrolledLoopBlocks.begin(), 815 UnrolledLoopBlocks.end(), Dest), 816 UnrolledLoopBlocks.end()); 817 } 818 } 819 } 820 821 // At this point, the code is well formed. We now simplify the unrolled loop, 822 // doing constant propagation and dead code elimination as we go. 823 simplifyLoopAfterUnroll(L, !CompletelyUnroll && (Count > 1 || Peeled), LI, SE, 824 DT, AC); 825 826 NumCompletelyUnrolled += CompletelyUnroll; 827 ++NumUnrolled; 828 829 Loop *OuterL = L->getParentLoop(); 830 // Update LoopInfo if the loop is completely removed. 831 if (CompletelyUnroll) 832 LI->erase(L); 833 834 // After complete unrolling most of the blocks should be contained in OuterL. 835 // However, some of them might happen to be out of OuterL (e.g. if they 836 // precede a loop exit). In this case we might need to insert PHI nodes in 837 // order to preserve LCSSA form. 838 // We don't need to check this if we already know that we need to fix LCSSA 839 // form. 840 // TODO: For now we just recompute LCSSA for the outer loop in this case, but 841 // it should be possible to fix it in-place. 842 if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA) 843 NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI); 844 845 // If we have a pass and a DominatorTree we should re-simplify impacted loops 846 // to ensure subsequent analyses can rely on this form. We want to simplify 847 // at least one layer outside of the loop that was unrolled so that any 848 // changes to the parent loop exposed by the unrolling are considered. 849 if (DT) { 850 if (OuterL) { 851 // OuterL includes all loops for which we can break loop-simplify, so 852 // it's sufficient to simplify only it (it'll recursively simplify inner 853 // loops too). 854 if (NeedToFixLCSSA) { 855 // LCSSA must be performed on the outermost affected loop. The unrolled 856 // loop's last loop latch is guaranteed to be in the outermost loop 857 // after LoopInfo's been updated by LoopInfo::erase. 858 Loop *LatchLoop = LI->getLoopFor(Latches.back()); 859 Loop *FixLCSSALoop = OuterL; 860 if (!FixLCSSALoop->contains(LatchLoop)) 861 while (FixLCSSALoop->getParentLoop() != LatchLoop) 862 FixLCSSALoop = FixLCSSALoop->getParentLoop(); 863 864 formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE); 865 } else if (PreserveLCSSA) { 866 assert(OuterL->isLCSSAForm(*DT) && 867 "Loops should be in LCSSA form after loop-unroll."); 868 } 869 870 // TODO: That potentially might be compile-time expensive. We should try 871 // to fix the loop-simplified form incrementally. 872 simplifyLoop(OuterL, DT, LI, SE, AC, PreserveLCSSA); 873 } else { 874 // Simplify loops for which we might've broken loop-simplify form. 875 for (Loop *SubLoop : LoopsToSimplify) 876 simplifyLoop(SubLoop, DT, LI, SE, AC, PreserveLCSSA); 877 } 878 } 879 880 return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled 881 : LoopUnrollResult::PartiallyUnrolled; 882 } 883 884 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata 885 /// node with the given name (for example, "llvm.loop.unroll.count"). If no 886 /// such metadata node exists, then nullptr is returned. 887 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) { 888 // First operand should refer to the loop id itself. 889 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 890 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 891 892 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 893 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 894 if (!MD) 895 continue; 896 897 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 898 if (!S) 899 continue; 900 901 if (Name.equals(S->getString())) 902 return MD; 903 } 904 return nullptr; 905 } 906