1 //===-- UnrollLoopRuntime.cpp - Runtime 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 for loops with run-time 10 // trip counts. See LoopUnroll.cpp for unrolling loops with compile-time 11 // trip counts. 12 // 13 // The functions in this file are used to generate extra code when the 14 // run-time trip count modulo the unroll factor is not 0. When this is the 15 // case, we need to generate code to execute these 'left over' iterations. 16 // 17 // The current strategy generates an if-then-else sequence prior to the 18 // unrolled loop to execute the 'left over' iterations before or after the 19 // unrolled loop. 20 // 21 //===----------------------------------------------------------------------===// 22 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/Analysis/DomTreeUpdater.h" 25 #include "llvm/Analysis/InstructionSimplify.h" 26 #include "llvm/Analysis/LoopIterator.h" 27 #include "llvm/Analysis/ScalarEvolution.h" 28 #include "llvm/Analysis/ValueTracking.h" 29 #include "llvm/IR/BasicBlock.h" 30 #include "llvm/IR/Dominators.h" 31 #include "llvm/IR/MDBuilder.h" 32 #include "llvm/IR/Module.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Support/raw_ostream.h" 36 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 37 #include "llvm/Transforms/Utils/Cloning.h" 38 #include "llvm/Transforms/Utils/Local.h" 39 #include "llvm/Transforms/Utils/LoopUtils.h" 40 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 41 #include "llvm/Transforms/Utils/UnrollLoop.h" 42 #include <algorithm> 43 44 using namespace llvm; 45 46 #define DEBUG_TYPE "loop-unroll" 47 48 STATISTIC(NumRuntimeUnrolled, 49 "Number of loops unrolled with run-time trip counts"); 50 static cl::opt<bool> UnrollRuntimeMultiExit( 51 "unroll-runtime-multi-exit", cl::init(false), cl::Hidden, 52 cl::desc("Allow runtime unrolling for loops with multiple exits, when " 53 "epilog is generated")); 54 static cl::opt<bool> UnrollRuntimeOtherExitPredictable( 55 "unroll-runtime-other-exit-predictable", cl::init(false), cl::Hidden, 56 cl::desc("Assume the non latch exit block to be predictable")); 57 58 /// Connect the unrolling prolog code to the original loop. 59 /// The unrolling prolog code contains code to execute the 60 /// 'extra' iterations if the run-time trip count modulo the 61 /// unroll count is non-zero. 62 /// 63 /// This function performs the following: 64 /// - Create PHI nodes at prolog end block to combine values 65 /// that exit the prolog code and jump around the prolog. 66 /// - Add a PHI operand to a PHI node at the loop exit block 67 /// for values that exit the prolog and go around the loop. 68 /// - Branch around the original loop if the trip count is less 69 /// than the unroll factor. 70 /// 71 static void ConnectProlog(Loop *L, Value *BECount, unsigned Count, 72 BasicBlock *PrologExit, 73 BasicBlock *OriginalLoopLatchExit, 74 BasicBlock *PreHeader, BasicBlock *NewPreHeader, 75 ValueToValueMapTy &VMap, DominatorTree *DT, 76 LoopInfo *LI, bool PreserveLCSSA, 77 ScalarEvolution &SE) { 78 // Loop structure should be the following: 79 // Preheader 80 // PrologHeader 81 // ... 82 // PrologLatch 83 // PrologExit 84 // NewPreheader 85 // Header 86 // ... 87 // Latch 88 // LatchExit 89 BasicBlock *Latch = L->getLoopLatch(); 90 assert(Latch && "Loop must have a latch"); 91 BasicBlock *PrologLatch = cast<BasicBlock>(VMap[Latch]); 92 93 // Create a PHI node for each outgoing value from the original loop 94 // (which means it is an outgoing value from the prolog code too). 95 // The new PHI node is inserted in the prolog end basic block. 96 // The new PHI node value is added as an operand of a PHI node in either 97 // the loop header or the loop exit block. 98 for (BasicBlock *Succ : successors(Latch)) { 99 for (PHINode &PN : Succ->phis()) { 100 // Add a new PHI node to the prolog end block and add the 101 // appropriate incoming values. 102 // TODO: This code assumes that the PrologExit (or the LatchExit block for 103 // prolog loop) contains only one predecessor from the loop, i.e. the 104 // PrologLatch. When supporting multiple-exiting block loops, we can have 105 // two or more blocks that have the LatchExit as the target in the 106 // original loop. 107 PHINode *NewPN = PHINode::Create(PN.getType(), 2, PN.getName() + ".unr", 108 PrologExit->getFirstNonPHI()); 109 // Adding a value to the new PHI node from the original loop preheader. 110 // This is the value that skips all the prolog code. 111 if (L->contains(&PN)) { 112 // Succ is loop header. 113 NewPN->addIncoming(PN.getIncomingValueForBlock(NewPreHeader), 114 PreHeader); 115 } else { 116 // Succ is LatchExit. 117 NewPN->addIncoming(UndefValue::get(PN.getType()), PreHeader); 118 } 119 120 Value *V = PN.getIncomingValueForBlock(Latch); 121 if (Instruction *I = dyn_cast<Instruction>(V)) { 122 if (L->contains(I)) { 123 V = VMap.lookup(I); 124 } 125 } 126 // Adding a value to the new PHI node from the last prolog block 127 // that was created. 128 NewPN->addIncoming(V, PrologLatch); 129 130 // Update the existing PHI node operand with the value from the 131 // new PHI node. How this is done depends on if the existing 132 // PHI node is in the original loop block, or the exit block. 133 if (L->contains(&PN)) 134 PN.setIncomingValueForBlock(NewPreHeader, NewPN); 135 else 136 PN.addIncoming(NewPN, PrologExit); 137 SE.forgetValue(&PN); 138 } 139 } 140 141 // Make sure that created prolog loop is in simplified form 142 SmallVector<BasicBlock *, 4> PrologExitPreds; 143 Loop *PrologLoop = LI->getLoopFor(PrologLatch); 144 if (PrologLoop) { 145 for (BasicBlock *PredBB : predecessors(PrologExit)) 146 if (PrologLoop->contains(PredBB)) 147 PrologExitPreds.push_back(PredBB); 148 149 SplitBlockPredecessors(PrologExit, PrologExitPreds, ".unr-lcssa", DT, LI, 150 nullptr, PreserveLCSSA); 151 } 152 153 // Create a branch around the original loop, which is taken if there are no 154 // iterations remaining to be executed after running the prologue. 155 Instruction *InsertPt = PrologExit->getTerminator(); 156 IRBuilder<> B(InsertPt); 157 158 assert(Count != 0 && "nonsensical Count!"); 159 160 // If BECount <u (Count - 1) then (BECount + 1) % Count == (BECount + 1) 161 // This means %xtraiter is (BECount + 1) and all of the iterations of this 162 // loop were executed by the prologue. Note that if BECount <u (Count - 1) 163 // then (BECount + 1) cannot unsigned-overflow. 164 Value *BrLoopExit = 165 B.CreateICmpULT(BECount, ConstantInt::get(BECount->getType(), Count - 1)); 166 // Split the exit to maintain loop canonicalization guarantees 167 SmallVector<BasicBlock *, 4> Preds(predecessors(OriginalLoopLatchExit)); 168 SplitBlockPredecessors(OriginalLoopLatchExit, Preds, ".unr-lcssa", DT, LI, 169 nullptr, PreserveLCSSA); 170 // Add the branch to the exit block (around the unrolled loop) 171 B.CreateCondBr(BrLoopExit, OriginalLoopLatchExit, NewPreHeader); 172 InsertPt->eraseFromParent(); 173 if (DT) { 174 auto *NewDom = DT->findNearestCommonDominator(OriginalLoopLatchExit, 175 PrologExit); 176 DT->changeImmediateDominator(OriginalLoopLatchExit, NewDom); 177 } 178 } 179 180 /// Connect the unrolling epilog code to the original loop. 181 /// The unrolling epilog code contains code to execute the 182 /// 'extra' iterations if the run-time trip count modulo the 183 /// unroll count is non-zero. 184 /// 185 /// This function performs the following: 186 /// - Update PHI nodes at the unrolling loop exit and epilog loop exit 187 /// - Create PHI nodes at the unrolling loop exit to combine 188 /// values that exit the unrolling loop code and jump around it. 189 /// - Update PHI operands in the epilog loop by the new PHI nodes 190 /// - Branch around the epilog loop if extra iters (ModVal) is zero. 191 /// 192 static void ConnectEpilog(Loop *L, Value *ModVal, BasicBlock *NewExit, 193 BasicBlock *Exit, BasicBlock *PreHeader, 194 BasicBlock *EpilogPreHeader, BasicBlock *NewPreHeader, 195 ValueToValueMapTy &VMap, DominatorTree *DT, 196 LoopInfo *LI, bool PreserveLCSSA, 197 ScalarEvolution &SE) { 198 BasicBlock *Latch = L->getLoopLatch(); 199 assert(Latch && "Loop must have a latch"); 200 BasicBlock *EpilogLatch = cast<BasicBlock>(VMap[Latch]); 201 202 // Loop structure should be the following: 203 // 204 // PreHeader 205 // NewPreHeader 206 // Header 207 // ... 208 // Latch 209 // NewExit (PN) 210 // EpilogPreHeader 211 // EpilogHeader 212 // ... 213 // EpilogLatch 214 // Exit (EpilogPN) 215 216 // Update PHI nodes at NewExit and Exit. 217 for (PHINode &PN : NewExit->phis()) { 218 // PN should be used in another PHI located in Exit block as 219 // Exit was split by SplitBlockPredecessors into Exit and NewExit 220 // Basicaly it should look like: 221 // NewExit: 222 // PN = PHI [I, Latch] 223 // ... 224 // Exit: 225 // EpilogPN = PHI [PN, EpilogPreHeader], [X, Exit2], [Y, Exit2.epil] 226 // 227 // Exits from non-latch blocks point to the original exit block and the 228 // epilogue edges have already been added. 229 // 230 // There is EpilogPreHeader incoming block instead of NewExit as 231 // NewExit was spilt 1 more time to get EpilogPreHeader. 232 assert(PN.hasOneUse() && "The phi should have 1 use"); 233 PHINode *EpilogPN = cast<PHINode>(PN.use_begin()->getUser()); 234 assert(EpilogPN->getParent() == Exit && "EpilogPN should be in Exit block"); 235 236 // Add incoming PreHeader from branch around the Loop 237 PN.addIncoming(UndefValue::get(PN.getType()), PreHeader); 238 SE.forgetValue(&PN); 239 240 Value *V = PN.getIncomingValueForBlock(Latch); 241 Instruction *I = dyn_cast<Instruction>(V); 242 if (I && L->contains(I)) 243 // If value comes from an instruction in the loop add VMap value. 244 V = VMap.lookup(I); 245 // For the instruction out of the loop, constant or undefined value 246 // insert value itself. 247 EpilogPN->addIncoming(V, EpilogLatch); 248 249 assert(EpilogPN->getBasicBlockIndex(EpilogPreHeader) >= 0 && 250 "EpilogPN should have EpilogPreHeader incoming block"); 251 // Change EpilogPreHeader incoming block to NewExit. 252 EpilogPN->setIncomingBlock(EpilogPN->getBasicBlockIndex(EpilogPreHeader), 253 NewExit); 254 // Now PHIs should look like: 255 // NewExit: 256 // PN = PHI [I, Latch], [undef, PreHeader] 257 // ... 258 // Exit: 259 // EpilogPN = PHI [PN, NewExit], [VMap[I], EpilogLatch] 260 } 261 262 // Create PHI nodes at NewExit (from the unrolling loop Latch and PreHeader). 263 // Update corresponding PHI nodes in epilog loop. 264 for (BasicBlock *Succ : successors(Latch)) { 265 // Skip this as we already updated phis in exit blocks. 266 if (!L->contains(Succ)) 267 continue; 268 for (PHINode &PN : Succ->phis()) { 269 // Add new PHI nodes to the loop exit block and update epilog 270 // PHIs with the new PHI values. 271 PHINode *NewPN = PHINode::Create(PN.getType(), 2, PN.getName() + ".unr", 272 NewExit->getFirstNonPHI()); 273 // Adding a value to the new PHI node from the unrolling loop preheader. 274 NewPN->addIncoming(PN.getIncomingValueForBlock(NewPreHeader), PreHeader); 275 // Adding a value to the new PHI node from the unrolling loop latch. 276 NewPN->addIncoming(PN.getIncomingValueForBlock(Latch), Latch); 277 278 // Update the existing PHI node operand with the value from the new PHI 279 // node. Corresponding instruction in epilog loop should be PHI. 280 PHINode *VPN = cast<PHINode>(VMap[&PN]); 281 VPN->setIncomingValueForBlock(EpilogPreHeader, NewPN); 282 } 283 } 284 285 Instruction *InsertPt = NewExit->getTerminator(); 286 IRBuilder<> B(InsertPt); 287 Value *BrLoopExit = B.CreateIsNotNull(ModVal, "lcmp.mod"); 288 assert(Exit && "Loop must have a single exit block only"); 289 // Split the epilogue exit to maintain loop canonicalization guarantees 290 SmallVector<BasicBlock*, 4> Preds(predecessors(Exit)); 291 SplitBlockPredecessors(Exit, Preds, ".epilog-lcssa", DT, LI, nullptr, 292 PreserveLCSSA); 293 // Add the branch to the exit block (around the unrolling loop) 294 B.CreateCondBr(BrLoopExit, EpilogPreHeader, Exit); 295 InsertPt->eraseFromParent(); 296 if (DT) { 297 auto *NewDom = DT->findNearestCommonDominator(Exit, NewExit); 298 DT->changeImmediateDominator(Exit, NewDom); 299 } 300 301 // Split the main loop exit to maintain canonicalization guarantees. 302 SmallVector<BasicBlock*, 4> NewExitPreds{Latch}; 303 SplitBlockPredecessors(NewExit, NewExitPreds, ".loopexit", DT, LI, nullptr, 304 PreserveLCSSA); 305 } 306 307 /// Create a clone of the blocks in a loop and connect them together. A new 308 /// loop will be created including all cloned blocks, and the iterator of the 309 /// new loop switched to count NewIter down to 0. 310 /// The cloned blocks should be inserted between InsertTop and InsertBot. 311 /// InsertTop should be new preheader, InsertBot new loop exit. 312 /// Returns the new cloned loop that is created. 313 static Loop * 314 CloneLoopBlocks(Loop *L, Value *NewIter, const bool UseEpilogRemainder, 315 const bool UnrollRemainder, 316 BasicBlock *InsertTop, 317 BasicBlock *InsertBot, BasicBlock *Preheader, 318 std::vector<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks, 319 ValueToValueMapTy &VMap, DominatorTree *DT, LoopInfo *LI) { 320 StringRef suffix = UseEpilogRemainder ? "epil" : "prol"; 321 BasicBlock *Header = L->getHeader(); 322 BasicBlock *Latch = L->getLoopLatch(); 323 Function *F = Header->getParent(); 324 LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO(); 325 LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO(); 326 Loop *ParentLoop = L->getParentLoop(); 327 NewLoopsMap NewLoops; 328 NewLoops[ParentLoop] = ParentLoop; 329 330 // For each block in the original loop, create a new copy, 331 // and update the value map with the newly created values. 332 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { 333 BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, "." + suffix, F); 334 NewBlocks.push_back(NewBB); 335 336 addClonedBlockToLoopInfo(*BB, NewBB, LI, NewLoops); 337 338 VMap[*BB] = NewBB; 339 if (Header == *BB) { 340 // For the first block, add a CFG connection to this newly 341 // created block. 342 InsertTop->getTerminator()->setSuccessor(0, NewBB); 343 } 344 345 if (DT) { 346 if (Header == *BB) { 347 // The header is dominated by the preheader. 348 DT->addNewBlock(NewBB, InsertTop); 349 } else { 350 // Copy information from original loop to unrolled loop. 351 BasicBlock *IDomBB = DT->getNode(*BB)->getIDom()->getBlock(); 352 DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDomBB])); 353 } 354 } 355 356 if (Latch == *BB) { 357 // For the last block, create a loop back to cloned head. 358 VMap.erase((*BB)->getTerminator()); 359 // Use an incrementing IV. Pre-incr/post-incr is backedge/trip count. 360 // Subtle: NewIter can be 0 if we wrapped when computing the trip count, 361 // thus we must compare the post-increment (wrapping) value. 362 BasicBlock *FirstLoopBB = cast<BasicBlock>(VMap[Header]); 363 BranchInst *LatchBR = cast<BranchInst>(NewBB->getTerminator()); 364 IRBuilder<> Builder(LatchBR); 365 PHINode *NewIdx = PHINode::Create(NewIter->getType(), 2, 366 suffix + ".iter", 367 FirstLoopBB->getFirstNonPHI()); 368 auto *Zero = ConstantInt::get(NewIdx->getType(), 0); 369 auto *One = ConstantInt::get(NewIdx->getType(), 1); 370 Value *IdxNext = Builder.CreateAdd(NewIdx, One, NewIdx->getName() + ".next"); 371 Value *IdxCmp = Builder.CreateICmpNE(IdxNext, NewIter, NewIdx->getName() + ".cmp"); 372 Builder.CreateCondBr(IdxCmp, FirstLoopBB, InsertBot); 373 NewIdx->addIncoming(Zero, InsertTop); 374 NewIdx->addIncoming(IdxNext, NewBB); 375 LatchBR->eraseFromParent(); 376 } 377 } 378 379 // Change the incoming values to the ones defined in the preheader or 380 // cloned loop. 381 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 382 PHINode *NewPHI = cast<PHINode>(VMap[&*I]); 383 unsigned idx = NewPHI->getBasicBlockIndex(Preheader); 384 NewPHI->setIncomingBlock(idx, InsertTop); 385 BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]); 386 idx = NewPHI->getBasicBlockIndex(Latch); 387 Value *InVal = NewPHI->getIncomingValue(idx); 388 NewPHI->setIncomingBlock(idx, NewLatch); 389 if (Value *V = VMap.lookup(InVal)) 390 NewPHI->setIncomingValue(idx, V); 391 } 392 393 Loop *NewLoop = NewLoops[L]; 394 assert(NewLoop && "L should have been cloned"); 395 MDNode *LoopID = NewLoop->getLoopID(); 396 397 // Only add loop metadata if the loop is not going to be completely 398 // unrolled. 399 if (UnrollRemainder) 400 return NewLoop; 401 402 Optional<MDNode *> NewLoopID = makeFollowupLoopID( 403 LoopID, {LLVMLoopUnrollFollowupAll, LLVMLoopUnrollFollowupRemainder}); 404 if (NewLoopID) { 405 NewLoop->setLoopID(NewLoopID.value()); 406 407 // Do not setLoopAlreadyUnrolled if loop attributes have been defined 408 // explicitly. 409 return NewLoop; 410 } 411 412 // Add unroll disable metadata to disable future unrolling for this loop. 413 NewLoop->setLoopAlreadyUnrolled(); 414 return NewLoop; 415 } 416 417 /// Returns true if we can profitably unroll the multi-exit loop L. Currently, 418 /// we return true only if UnrollRuntimeMultiExit is set to true. 419 static bool canProfitablyUnrollMultiExitLoop( 420 Loop *L, SmallVectorImpl<BasicBlock *> &OtherExits, BasicBlock *LatchExit, 421 bool UseEpilogRemainder) { 422 423 // Priority goes to UnrollRuntimeMultiExit if it's supplied. 424 if (UnrollRuntimeMultiExit.getNumOccurrences()) 425 return UnrollRuntimeMultiExit; 426 427 // The main pain point with multi-exit loop unrolling is that once unrolled, 428 // we will not be able to merge all blocks into a straight line code. 429 // There are branches within the unrolled loop that go to the OtherExits. 430 // The second point is the increase in code size, but this is true 431 // irrespective of multiple exits. 432 433 // Note: Both the heuristics below are coarse grained. We are essentially 434 // enabling unrolling of loops that have a single side exit other than the 435 // normal LatchExit (i.e. exiting into a deoptimize block). 436 // The heuristics considered are: 437 // 1. low number of branches in the unrolled version. 438 // 2. high predictability of these extra branches. 439 // We avoid unrolling loops that have more than two exiting blocks. This 440 // limits the total number of branches in the unrolled loop to be atmost 441 // the unroll factor (since one of the exiting blocks is the latch block). 442 SmallVector<BasicBlock*, 4> ExitingBlocks; 443 L->getExitingBlocks(ExitingBlocks); 444 if (ExitingBlocks.size() > 2) 445 return false; 446 447 // Allow unrolling of loops with no non latch exit blocks. 448 if (OtherExits.size() == 0) 449 return true; 450 451 // The second heuristic is that L has one exit other than the latchexit and 452 // that exit is a deoptimize block. We know that deoptimize blocks are rarely 453 // taken, which also implies the branch leading to the deoptimize block is 454 // highly predictable. When UnrollRuntimeOtherExitPredictable is specified, we 455 // assume the other exit branch is predictable even if it has no deoptimize 456 // call. 457 return (OtherExits.size() == 1 && 458 (UnrollRuntimeOtherExitPredictable || 459 OtherExits[0]->getTerminatingDeoptimizeCall())); 460 // TODO: These can be fine-tuned further to consider code size or deopt states 461 // that are captured by the deoptimize exit block. 462 // Also, we can extend this to support more cases, if we actually 463 // know of kinds of multiexit loops that would benefit from unrolling. 464 } 465 466 // Assign the maximum possible trip count as the back edge weight for the 467 // remainder loop if the original loop comes with a branch weight. 468 static void updateLatchBranchWeightsForRemainderLoop(Loop *OrigLoop, 469 Loop *RemainderLoop, 470 uint64_t UnrollFactor) { 471 uint64_t TrueWeight, FalseWeight; 472 BranchInst *LatchBR = 473 cast<BranchInst>(OrigLoop->getLoopLatch()->getTerminator()); 474 if (!LatchBR->extractProfMetadata(TrueWeight, FalseWeight)) 475 return; 476 uint64_t ExitWeight = LatchBR->getSuccessor(0) == OrigLoop->getHeader() 477 ? FalseWeight 478 : TrueWeight; 479 assert(UnrollFactor > 1); 480 uint64_t BackEdgeWeight = (UnrollFactor - 1) * ExitWeight; 481 BasicBlock *Header = RemainderLoop->getHeader(); 482 BasicBlock *Latch = RemainderLoop->getLoopLatch(); 483 auto *RemainderLatchBR = cast<BranchInst>(Latch->getTerminator()); 484 unsigned HeaderIdx = (RemainderLatchBR->getSuccessor(0) == Header ? 0 : 1); 485 MDBuilder MDB(RemainderLatchBR->getContext()); 486 MDNode *WeightNode = 487 HeaderIdx ? MDB.createBranchWeights(ExitWeight, BackEdgeWeight) 488 : MDB.createBranchWeights(BackEdgeWeight, ExitWeight); 489 RemainderLatchBR->setMetadata(LLVMContext::MD_prof, WeightNode); 490 } 491 492 /// Calculate ModVal = (BECount + 1) % Count on the abstract integer domain 493 /// accounting for the possibility of unsigned overflow in the 2s complement 494 /// domain. Preconditions: 495 /// 1) TripCount = BECount + 1 (allowing overflow) 496 /// 2) Log2(Count) <= BitWidth(BECount) 497 static Value *CreateTripRemainder(IRBuilder<> &B, Value *BECount, 498 Value *TripCount, unsigned Count) { 499 // Note that TripCount is BECount + 1. 500 if (isPowerOf2_32(Count)) 501 // If the expression is zero, then either: 502 // 1. There are no iterations to be run in the prolog/epilog loop. 503 // OR 504 // 2. The addition computing TripCount overflowed. 505 // 506 // If (2) is true, we know that TripCount really is (1 << BEWidth) and so 507 // the number of iterations that remain to be run in the original loop is a 508 // multiple Count == (1 << Log2(Count)) because Log2(Count) <= BEWidth (a 509 // precondition of this method). 510 return B.CreateAnd(TripCount, Count - 1, "xtraiter"); 511 512 // As (BECount + 1) can potentially unsigned overflow we count 513 // (BECount % Count) + 1 which is overflow safe as BECount % Count < Count. 514 Constant *CountC = ConstantInt::get(BECount->getType(), Count); 515 Value *ModValTmp = B.CreateURem(BECount, CountC); 516 Value *ModValAdd = B.CreateAdd(ModValTmp, 517 ConstantInt::get(ModValTmp->getType(), 1)); 518 // At that point (BECount % Count) + 1 could be equal to Count. 519 // To handle this case we need to take mod by Count one more time. 520 return B.CreateURem(ModValAdd, CountC, "xtraiter"); 521 } 522 523 524 /// Insert code in the prolog/epilog code when unrolling a loop with a 525 /// run-time trip-count. 526 /// 527 /// This method assumes that the loop unroll factor is total number 528 /// of loop bodies in the loop after unrolling. (Some folks refer 529 /// to the unroll factor as the number of *extra* copies added). 530 /// We assume also that the loop unroll factor is a power-of-two. So, after 531 /// unrolling the loop, the number of loop bodies executed is 2, 532 /// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch 533 /// instruction in SimplifyCFG.cpp. Then, the backend decides how code for 534 /// the switch instruction is generated. 535 /// 536 /// ***Prolog case*** 537 /// extraiters = tripcount % loopfactor 538 /// if (extraiters == 0) jump Loop: 539 /// else jump Prol: 540 /// Prol: LoopBody; 541 /// extraiters -= 1 // Omitted if unroll factor is 2. 542 /// if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2. 543 /// if (tripcount < loopfactor) jump End: 544 /// Loop: 545 /// ... 546 /// End: 547 /// 548 /// ***Epilog case*** 549 /// extraiters = tripcount % loopfactor 550 /// if (tripcount < loopfactor) jump LoopExit: 551 /// unroll_iters = tripcount - extraiters 552 /// Loop: LoopBody; (executes unroll_iter times); 553 /// unroll_iter -= 1 554 /// if (unroll_iter != 0) jump Loop: 555 /// LoopExit: 556 /// if (extraiters == 0) jump EpilExit: 557 /// Epil: LoopBody; (executes extraiters times) 558 /// extraiters -= 1 // Omitted if unroll factor is 2. 559 /// if (extraiters != 0) jump Epil: // Omitted if unroll factor is 2. 560 /// EpilExit: 561 562 bool llvm::UnrollRuntimeLoopRemainder( 563 Loop *L, unsigned Count, bool AllowExpensiveTripCount, 564 bool UseEpilogRemainder, bool UnrollRemainder, bool ForgetAllSCEV, 565 LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC, 566 const TargetTransformInfo *TTI, bool PreserveLCSSA, Loop **ResultLoop) { 567 LLVM_DEBUG(dbgs() << "Trying runtime unrolling on Loop: \n"); 568 LLVM_DEBUG(L->dump()); 569 LLVM_DEBUG(UseEpilogRemainder ? dbgs() << "Using epilog remainder.\n" 570 : dbgs() << "Using prolog remainder.\n"); 571 572 // Make sure the loop is in canonical form. 573 if (!L->isLoopSimplifyForm()) { 574 LLVM_DEBUG(dbgs() << "Not in simplify form!\n"); 575 return false; 576 } 577 578 // Guaranteed by LoopSimplifyForm. 579 BasicBlock *Latch = L->getLoopLatch(); 580 BasicBlock *Header = L->getHeader(); 581 582 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator()); 583 584 if (!LatchBR || LatchBR->isUnconditional()) { 585 // The loop-rotate pass can be helpful to avoid this in many cases. 586 LLVM_DEBUG( 587 dbgs() 588 << "Loop latch not terminated by a conditional branch.\n"); 589 return false; 590 } 591 592 unsigned ExitIndex = LatchBR->getSuccessor(0) == Header ? 1 : 0; 593 BasicBlock *LatchExit = LatchBR->getSuccessor(ExitIndex); 594 595 if (L->contains(LatchExit)) { 596 // Cloning the loop basic blocks (`CloneLoopBlocks`) requires that one of the 597 // targets of the Latch be an exit block out of the loop. 598 LLVM_DEBUG( 599 dbgs() 600 << "One of the loop latch successors must be the exit block.\n"); 601 return false; 602 } 603 604 // These are exit blocks other than the target of the latch exiting block. 605 SmallVector<BasicBlock *, 4> OtherExits; 606 L->getUniqueNonLatchExitBlocks(OtherExits); 607 // Support only single exit and exiting block unless multi-exit loop 608 // unrolling is enabled. 609 if (!L->getExitingBlock() || OtherExits.size()) { 610 // We rely on LCSSA form being preserved when the exit blocks are transformed. 611 // (Note that only an off-by-default mode of the old PM disables PreserveLCCA.) 612 if (!PreserveLCSSA) 613 return false; 614 615 if (!canProfitablyUnrollMultiExitLoop(L, OtherExits, LatchExit, 616 UseEpilogRemainder)) { 617 LLVM_DEBUG( 618 dbgs() 619 << "Multiple exit/exiting blocks in loop and multi-exit unrolling not " 620 "enabled!\n"); 621 return false; 622 } 623 } 624 // Use Scalar Evolution to compute the trip count. This allows more loops to 625 // be unrolled than relying on induction var simplification. 626 if (!SE) 627 return false; 628 629 // Only unroll loops with a computable trip count. 630 // We calculate the backedge count by using getExitCount on the Latch block, 631 // which is proven to be the only exiting block in this loop. This is same as 632 // calculating getBackedgeTakenCount on the loop (which computes SCEV for all 633 // exiting blocks). 634 const SCEV *BECountSC = SE->getExitCount(L, Latch); 635 if (isa<SCEVCouldNotCompute>(BECountSC)) { 636 LLVM_DEBUG(dbgs() << "Could not compute exit block SCEV\n"); 637 return false; 638 } 639 640 unsigned BEWidth = cast<IntegerType>(BECountSC->getType())->getBitWidth(); 641 642 // Add 1 since the backedge count doesn't include the first loop iteration. 643 // (Note that overflow can occur, this is handled explicitly below) 644 const SCEV *TripCountSC = 645 SE->getAddExpr(BECountSC, SE->getConstant(BECountSC->getType(), 1)); 646 if (isa<SCEVCouldNotCompute>(TripCountSC)) { 647 LLVM_DEBUG(dbgs() << "Could not compute trip count SCEV.\n"); 648 return false; 649 } 650 651 BasicBlock *PreHeader = L->getLoopPreheader(); 652 BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator()); 653 const DataLayout &DL = Header->getModule()->getDataLayout(); 654 SCEVExpander Expander(*SE, DL, "loop-unroll"); 655 if (!AllowExpensiveTripCount && 656 Expander.isHighCostExpansion(TripCountSC, L, SCEVCheapExpansionBudget, 657 TTI, PreHeaderBR)) { 658 LLVM_DEBUG(dbgs() << "High cost for expanding trip count scev!\n"); 659 return false; 660 } 661 662 // This constraint lets us deal with an overflowing trip count easily; see the 663 // comment on ModVal below. 664 if (Log2_32(Count) > BEWidth) { 665 LLVM_DEBUG( 666 dbgs() 667 << "Count failed constraint on overflow trip count calculation.\n"); 668 return false; 669 } 670 671 // Loop structure is the following: 672 // 673 // PreHeader 674 // Header 675 // ... 676 // Latch 677 // LatchExit 678 679 BasicBlock *NewPreHeader; 680 BasicBlock *NewExit = nullptr; 681 BasicBlock *PrologExit = nullptr; 682 BasicBlock *EpilogPreHeader = nullptr; 683 BasicBlock *PrologPreHeader = nullptr; 684 685 if (UseEpilogRemainder) { 686 // If epilog remainder 687 // Split PreHeader to insert a branch around loop for unrolling. 688 NewPreHeader = SplitBlock(PreHeader, PreHeader->getTerminator(), DT, LI); 689 NewPreHeader->setName(PreHeader->getName() + ".new"); 690 // Split LatchExit to create phi nodes from branch above. 691 NewExit = SplitBlockPredecessors(LatchExit, {Latch}, ".unr-lcssa", DT, LI, 692 nullptr, PreserveLCSSA); 693 // NewExit gets its DebugLoc from LatchExit, which is not part of the 694 // original Loop. 695 // Fix this by setting Loop's DebugLoc to NewExit. 696 auto *NewExitTerminator = NewExit->getTerminator(); 697 NewExitTerminator->setDebugLoc(Header->getTerminator()->getDebugLoc()); 698 // Split NewExit to insert epilog remainder loop. 699 EpilogPreHeader = SplitBlock(NewExit, NewExitTerminator, DT, LI); 700 EpilogPreHeader->setName(Header->getName() + ".epil.preheader"); 701 702 // If the latch exits from multiple level of nested loops, then 703 // by assumption there must be another loop exit which branches to the 704 // outer loop and we must adjust the loop for the newly inserted blocks 705 // to account for the fact that our epilogue is still in the same outer 706 // loop. Note that this leaves loopinfo temporarily out of sync with the 707 // CFG until the actual epilogue loop is inserted. 708 if (auto *ParentL = L->getParentLoop()) 709 if (LI->getLoopFor(LatchExit) != ParentL) { 710 LI->removeBlock(NewExit); 711 ParentL->addBasicBlockToLoop(NewExit, *LI); 712 LI->removeBlock(EpilogPreHeader); 713 ParentL->addBasicBlockToLoop(EpilogPreHeader, *LI); 714 } 715 716 } else { 717 // If prolog remainder 718 // Split the original preheader twice to insert prolog remainder loop 719 PrologPreHeader = SplitEdge(PreHeader, Header, DT, LI); 720 PrologPreHeader->setName(Header->getName() + ".prol.preheader"); 721 PrologExit = SplitBlock(PrologPreHeader, PrologPreHeader->getTerminator(), 722 DT, LI); 723 PrologExit->setName(Header->getName() + ".prol.loopexit"); 724 // Split PrologExit to get NewPreHeader. 725 NewPreHeader = SplitBlock(PrologExit, PrologExit->getTerminator(), DT, LI); 726 NewPreHeader->setName(PreHeader->getName() + ".new"); 727 } 728 // Loop structure should be the following: 729 // Epilog Prolog 730 // 731 // PreHeader PreHeader 732 // *NewPreHeader *PrologPreHeader 733 // Header *PrologExit 734 // ... *NewPreHeader 735 // Latch Header 736 // *NewExit ... 737 // *EpilogPreHeader Latch 738 // LatchExit LatchExit 739 740 // Calculate conditions for branch around loop for unrolling 741 // in epilog case and around prolog remainder loop in prolog case. 742 // Compute the number of extra iterations required, which is: 743 // extra iterations = run-time trip count % loop unroll factor 744 PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator()); 745 IRBuilder<> B(PreHeaderBR); 746 Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(), 747 PreHeaderBR); 748 Value *BECount; 749 // If there are other exits before the latch, that may cause the latch exit 750 // branch to never be executed, and the latch exit count may be poison. 751 // In this case, freeze the TripCount and base BECount on the frozen 752 // TripCount. We will introduce two branches using these values, and it's 753 // important that they see a consistent value (which would not be guaranteed 754 // if were frozen independently.) 755 if ((!OtherExits.empty() || !SE->loopHasNoAbnormalExits(L)) && 756 !isGuaranteedNotToBeUndefOrPoison(TripCount, AC, PreHeaderBR, DT)) { 757 TripCount = B.CreateFreeze(TripCount); 758 BECount = 759 B.CreateAdd(TripCount, ConstantInt::get(TripCount->getType(), -1)); 760 } else { 761 // If we don't need to freeze, use SCEVExpander for BECount as well, to 762 // allow slightly better value reuse. 763 BECount = 764 Expander.expandCodeFor(BECountSC, BECountSC->getType(), PreHeaderBR); 765 } 766 767 Value * const ModVal = CreateTripRemainder(B, BECount, TripCount, Count); 768 769 Value *BranchVal = 770 UseEpilogRemainder ? B.CreateICmpULT(BECount, 771 ConstantInt::get(BECount->getType(), 772 Count - 1)) : 773 B.CreateIsNotNull(ModVal, "lcmp.mod"); 774 BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader; 775 BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit; 776 // Branch to either remainder (extra iterations) loop or unrolling loop. 777 B.CreateCondBr(BranchVal, RemainderLoop, UnrollingLoop); 778 PreHeaderBR->eraseFromParent(); 779 if (DT) { 780 if (UseEpilogRemainder) 781 DT->changeImmediateDominator(NewExit, PreHeader); 782 else 783 DT->changeImmediateDominator(PrologExit, PreHeader); 784 } 785 Function *F = Header->getParent(); 786 // Get an ordered list of blocks in the loop to help with the ordering of the 787 // cloned blocks in the prolog/epilog code 788 LoopBlocksDFS LoopBlocks(L); 789 LoopBlocks.perform(LI); 790 791 // 792 // For each extra loop iteration, create a copy of the loop's basic blocks 793 // and generate a condition that branches to the copy depending on the 794 // number of 'left over' iterations. 795 // 796 std::vector<BasicBlock *> NewBlocks; 797 ValueToValueMapTy VMap; 798 799 // Clone all the basic blocks in the loop. If Count is 2, we don't clone 800 // the loop, otherwise we create a cloned loop to execute the extra 801 // iterations. This function adds the appropriate CFG connections. 802 BasicBlock *InsertBot = UseEpilogRemainder ? LatchExit : PrologExit; 803 BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader; 804 Loop *remainderLoop = CloneLoopBlocks( 805 L, ModVal, UseEpilogRemainder, UnrollRemainder, InsertTop, InsertBot, 806 NewPreHeader, NewBlocks, LoopBlocks, VMap, DT, LI); 807 808 // Assign the maximum possible trip count as the back edge weight for the 809 // remainder loop if the original loop comes with a branch weight. 810 if (remainderLoop && !UnrollRemainder) 811 updateLatchBranchWeightsForRemainderLoop(L, remainderLoop, Count); 812 813 // Insert the cloned blocks into the function. 814 F->getBasicBlockList().splice(InsertBot->getIterator(), 815 F->getBasicBlockList(), 816 NewBlocks[0]->getIterator(), 817 F->end()); 818 819 // Now the loop blocks are cloned and the other exiting blocks from the 820 // remainder are connected to the original Loop's exit blocks. The remaining 821 // work is to update the phi nodes in the original loop, and take in the 822 // values from the cloned region. 823 for (auto *BB : OtherExits) { 824 // Given we preserve LCSSA form, we know that the values used outside the 825 // loop will be used through these phi nodes at the exit blocks that are 826 // transformed below. 827 for (PHINode &PN : BB->phis()) { 828 unsigned oldNumOperands = PN.getNumIncomingValues(); 829 // Add the incoming values from the remainder code to the end of the phi 830 // node. 831 for (unsigned i = 0; i < oldNumOperands; i++){ 832 auto *PredBB =PN.getIncomingBlock(i); 833 if (PredBB == Latch) 834 // The latch exit is handled seperately, see connectX 835 continue; 836 if (!L->contains(PredBB)) 837 // Even if we had dedicated exits, the code above inserted an 838 // extra branch which can reach the latch exit. 839 continue; 840 841 auto *V = PN.getIncomingValue(i); 842 if (Instruction *I = dyn_cast<Instruction>(V)) 843 if (L->contains(I)) 844 V = VMap.lookup(I); 845 PN.addIncoming(V, cast<BasicBlock>(VMap[PredBB])); 846 } 847 } 848 #if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG) 849 for (BasicBlock *SuccBB : successors(BB)) { 850 assert(!(llvm::is_contained(OtherExits, SuccBB) || SuccBB == LatchExit) && 851 "Breaks the definition of dedicated exits!"); 852 } 853 #endif 854 } 855 856 // Update the immediate dominator of the exit blocks and blocks that are 857 // reachable from the exit blocks. This is needed because we now have paths 858 // from both the original loop and the remainder code reaching the exit 859 // blocks. While the IDom of these exit blocks were from the original loop, 860 // now the IDom is the preheader (which decides whether the original loop or 861 // remainder code should run). 862 if (DT && !L->getExitingBlock()) { 863 SmallVector<BasicBlock *, 16> ChildrenToUpdate; 864 // NB! We have to examine the dom children of all loop blocks, not just 865 // those which are the IDom of the exit blocks. This is because blocks 866 // reachable from the exit blocks can have their IDom as the nearest common 867 // dominator of the exit blocks. 868 for (auto *BB : L->blocks()) { 869 auto *DomNodeBB = DT->getNode(BB); 870 for (auto *DomChild : DomNodeBB->children()) { 871 auto *DomChildBB = DomChild->getBlock(); 872 if (!L->contains(LI->getLoopFor(DomChildBB))) 873 ChildrenToUpdate.push_back(DomChildBB); 874 } 875 } 876 for (auto *BB : ChildrenToUpdate) 877 DT->changeImmediateDominator(BB, PreHeader); 878 } 879 880 // Loop structure should be the following: 881 // Epilog Prolog 882 // 883 // PreHeader PreHeader 884 // NewPreHeader PrologPreHeader 885 // Header PrologHeader 886 // ... ... 887 // Latch PrologLatch 888 // NewExit PrologExit 889 // EpilogPreHeader NewPreHeader 890 // EpilogHeader Header 891 // ... ... 892 // EpilogLatch Latch 893 // LatchExit LatchExit 894 895 // Rewrite the cloned instruction operands to use the values created when the 896 // clone is created. 897 for (BasicBlock *BB : NewBlocks) { 898 for (Instruction &I : *BB) { 899 RemapInstruction(&I, VMap, 900 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 901 } 902 } 903 904 if (UseEpilogRemainder) { 905 // Connect the epilog code to the original loop and update the 906 // PHI functions. 907 ConnectEpilog(L, ModVal, NewExit, LatchExit, PreHeader, EpilogPreHeader, 908 NewPreHeader, VMap, DT, LI, PreserveLCSSA, *SE); 909 910 // Update counter in loop for unrolling. 911 // Use an incrementing IV. Pre-incr/post-incr is backedge/trip count. 912 // Subtle: TestVal can be 0 if we wrapped when computing the trip count, 913 // thus we must compare the post-increment (wrapping) value. 914 IRBuilder<> B2(NewPreHeader->getTerminator()); 915 Value *TestVal = B2.CreateSub(TripCount, ModVal, "unroll_iter"); 916 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator()); 917 PHINode *NewIdx = PHINode::Create(TestVal->getType(), 2, "niter", 918 Header->getFirstNonPHI()); 919 B2.SetInsertPoint(LatchBR); 920 auto *Zero = ConstantInt::get(NewIdx->getType(), 0); 921 auto *One = ConstantInt::get(NewIdx->getType(), 1); 922 Value *IdxNext = B2.CreateAdd(NewIdx, One, NewIdx->getName() + ".next"); 923 auto Pred = LatchBR->getSuccessor(0) == Header ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ; 924 Value *IdxCmp = B2.CreateICmp(Pred, IdxNext, TestVal, NewIdx->getName() + ".ncmp"); 925 NewIdx->addIncoming(Zero, NewPreHeader); 926 NewIdx->addIncoming(IdxNext, Latch); 927 LatchBR->setCondition(IdxCmp); 928 } else { 929 // Connect the prolog code to the original loop and update the 930 // PHI functions. 931 ConnectProlog(L, BECount, Count, PrologExit, LatchExit, PreHeader, 932 NewPreHeader, VMap, DT, LI, PreserveLCSSA, *SE); 933 } 934 935 // If this loop is nested, then the loop unroller changes the code in the any 936 // of its parent loops, so the Scalar Evolution pass needs to be run again. 937 SE->forgetTopmostLoop(L); 938 939 // Verify that the Dom Tree and Loop Info are correct. 940 #if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG) 941 if (DT) { 942 assert(DT->verify(DominatorTree::VerificationLevel::Full)); 943 LI->verify(*DT); 944 } 945 #endif 946 947 // For unroll factor 2 remainder loop will have 1 iteration. 948 if (Count == 2 && DT && LI && SE) { 949 // TODO: This code could probably be pulled out into a helper function 950 // (e.g. breakLoopBackedgeAndSimplify) and reused in loop-deletion. 951 BasicBlock *RemainderLatch = remainderLoop->getLoopLatch(); 952 assert(RemainderLatch); 953 SmallVector<BasicBlock*> RemainderBlocks(remainderLoop->getBlocks().begin(), 954 remainderLoop->getBlocks().end()); 955 breakLoopBackedge(remainderLoop, *DT, *SE, *LI, nullptr); 956 remainderLoop = nullptr; 957 958 // Simplify loop values after breaking the backedge 959 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 960 SmallVector<WeakTrackingVH, 16> DeadInsts; 961 for (BasicBlock *BB : RemainderBlocks) { 962 for (Instruction &Inst : llvm::make_early_inc_range(*BB)) { 963 if (Value *V = simplifyInstruction(&Inst, {DL, nullptr, DT, AC})) 964 if (LI->replacementPreservesLCSSAForm(&Inst, V)) 965 Inst.replaceAllUsesWith(V); 966 if (isInstructionTriviallyDead(&Inst)) 967 DeadInsts.emplace_back(&Inst); 968 } 969 // We can't do recursive deletion until we're done iterating, as we might 970 // have a phi which (potentially indirectly) uses instructions later in 971 // the block we're iterating through. 972 RecursivelyDeleteTriviallyDeadInstructions(DeadInsts); 973 } 974 975 // Merge latch into exit block. 976 auto *ExitBB = RemainderLatch->getSingleSuccessor(); 977 assert(ExitBB && "required after breaking cond br backedge"); 978 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 979 MergeBlockIntoPredecessor(ExitBB, &DTU, LI); 980 } 981 982 // Canonicalize to LoopSimplifyForm both original and remainder loops. We 983 // cannot rely on the LoopUnrollPass to do this because it only does 984 // canonicalization for parent/subloops and not the sibling loops. 985 if (OtherExits.size() > 0) { 986 // Generate dedicated exit blocks for the original loop, to preserve 987 // LoopSimplifyForm. 988 formDedicatedExitBlocks(L, DT, LI, nullptr, PreserveLCSSA); 989 // Generate dedicated exit blocks for the remainder loop if one exists, to 990 // preserve LoopSimplifyForm. 991 if (remainderLoop) 992 formDedicatedExitBlocks(remainderLoop, DT, LI, nullptr, PreserveLCSSA); 993 } 994 995 auto UnrollResult = LoopUnrollResult::Unmodified; 996 if (remainderLoop && UnrollRemainder) { 997 LLVM_DEBUG(dbgs() << "Unrolling remainder loop\n"); 998 UnrollResult = 999 UnrollLoop(remainderLoop, 1000 {/*Count*/ Count - 1, /*Force*/ false, /*Runtime*/ false, 1001 /*AllowExpensiveTripCount*/ false, 1002 /*UnrollRemainder*/ false, ForgetAllSCEV}, 1003 LI, SE, DT, AC, TTI, /*ORE*/ nullptr, PreserveLCSSA); 1004 } 1005 1006 if (ResultLoop && UnrollResult != LoopUnrollResult::FullyUnrolled) 1007 *ResultLoop = remainderLoop; 1008 NumRuntimeUnrolled++; 1009 return true; 1010 } 1011