1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===// 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 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 10 #include "llvm/ADT/DenseMap.h" 11 #include "llvm/ADT/STLExtras.h" 12 #include "llvm/ADT/Sequence.h" 13 #include "llvm/ADT/SetVector.h" 14 #include "llvm/ADT/SmallPtrSet.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/Statistic.h" 17 #include "llvm/ADT/Twine.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/BlockFrequencyInfo.h" 20 #include "llvm/Analysis/CFG.h" 21 #include "llvm/Analysis/CodeMetrics.h" 22 #include "llvm/Analysis/DomTreeUpdater.h" 23 #include "llvm/Analysis/GuardUtils.h" 24 #include "llvm/Analysis/LoopAnalysisManager.h" 25 #include "llvm/Analysis/LoopInfo.h" 26 #include "llvm/Analysis/LoopIterator.h" 27 #include "llvm/Analysis/MemorySSA.h" 28 #include "llvm/Analysis/MemorySSAUpdater.h" 29 #include "llvm/Analysis/MustExecute.h" 30 #include "llvm/Analysis/ProfileSummaryInfo.h" 31 #include "llvm/Analysis/ScalarEvolution.h" 32 #include "llvm/Analysis/TargetTransformInfo.h" 33 #include "llvm/Analysis/ValueTracking.h" 34 #include "llvm/IR/BasicBlock.h" 35 #include "llvm/IR/Constant.h" 36 #include "llvm/IR/Constants.h" 37 #include "llvm/IR/Dominators.h" 38 #include "llvm/IR/Function.h" 39 #include "llvm/IR/IRBuilder.h" 40 #include "llvm/IR/InstrTypes.h" 41 #include "llvm/IR/Instruction.h" 42 #include "llvm/IR/Instructions.h" 43 #include "llvm/IR/IntrinsicInst.h" 44 #include "llvm/IR/PatternMatch.h" 45 #include "llvm/IR/ProfDataUtils.h" 46 #include "llvm/IR/Use.h" 47 #include "llvm/IR/Value.h" 48 #include "llvm/Support/Casting.h" 49 #include "llvm/Support/CommandLine.h" 50 #include "llvm/Support/Debug.h" 51 #include "llvm/Support/ErrorHandling.h" 52 #include "llvm/Support/GenericDomTree.h" 53 #include "llvm/Support/InstructionCost.h" 54 #include "llvm/Support/raw_ostream.h" 55 #include "llvm/Transforms/Scalar/LoopPassManager.h" 56 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 57 #include "llvm/Transforms/Utils/Cloning.h" 58 #include "llvm/Transforms/Utils/Local.h" 59 #include "llvm/Transforms/Utils/LoopUtils.h" 60 #include "llvm/Transforms/Utils/ValueMapper.h" 61 #include <algorithm> 62 #include <cassert> 63 #include <iterator> 64 #include <numeric> 65 #include <optional> 66 #include <utility> 67 68 #define DEBUG_TYPE "simple-loop-unswitch" 69 70 using namespace llvm; 71 using namespace llvm::PatternMatch; 72 73 STATISTIC(NumBranches, "Number of branches unswitched"); 74 STATISTIC(NumSwitches, "Number of switches unswitched"); 75 STATISTIC(NumSelects, "Number of selects turned into branches for unswitching"); 76 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching"); 77 STATISTIC(NumTrivial, "Number of unswitches that are trivial"); 78 STATISTIC( 79 NumCostMultiplierSkipped, 80 "Number of unswitch candidates that had their cost multiplier skipped"); 81 STATISTIC(NumInvariantConditionsInjected, 82 "Number of invariant conditions injected and unswitched"); 83 84 static cl::opt<bool> EnableNonTrivialUnswitch( 85 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden, 86 cl::desc("Forcibly enables non-trivial loop unswitching rather than " 87 "following the configuration passed into the pass.")); 88 89 static cl::opt<int> 90 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, 91 cl::desc("The cost threshold for unswitching a loop.")); 92 93 static cl::opt<bool> EnableUnswitchCostMultiplier( 94 "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden, 95 cl::desc("Enable unswitch cost multiplier that prohibits exponential " 96 "explosion in nontrivial unswitch.")); 97 static cl::opt<int> UnswitchSiblingsToplevelDiv( 98 "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden, 99 cl::desc("Toplevel siblings divisor for cost multiplier.")); 100 static cl::opt<int> UnswitchNumInitialUnscaledCandidates( 101 "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden, 102 cl::desc("Number of unswitch candidates that are ignored when calculating " 103 "cost multiplier.")); 104 static cl::opt<bool> UnswitchGuards( 105 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden, 106 cl::desc("If enabled, simple loop unswitching will also consider " 107 "llvm.experimental.guard intrinsics as unswitch candidates.")); 108 static cl::opt<bool> DropNonTrivialImplicitNullChecks( 109 "simple-loop-unswitch-drop-non-trivial-implicit-null-checks", 110 cl::init(false), cl::Hidden, 111 cl::desc("If enabled, drop make.implicit metadata in unswitched implicit " 112 "null checks to save time analyzing if we can keep it.")); 113 static cl::opt<unsigned> 114 MSSAThreshold("simple-loop-unswitch-memoryssa-threshold", 115 cl::desc("Max number of memory uses to explore during " 116 "partial unswitching analysis"), 117 cl::init(100), cl::Hidden); 118 static cl::opt<bool> FreezeLoopUnswitchCond( 119 "freeze-loop-unswitch-cond", cl::init(true), cl::Hidden, 120 cl::desc("If enabled, the freeze instruction will be added to condition " 121 "of loop unswitch to prevent miscompilation.")); 122 123 static cl::opt<bool> InjectInvariantConditions( 124 "simple-loop-unswitch-inject-invariant-conditions", cl::Hidden, 125 cl::desc("Whether we should inject new invariants and unswitch them to " 126 "eliminate some existing (non-invariant) conditions."), 127 cl::init(true)); 128 129 static cl::opt<unsigned> InjectInvariantConditionHotnesThreshold( 130 "simple-loop-unswitch-inject-invariant-condition-hotness-threshold", 131 cl::Hidden, cl::desc("Only try to inject loop invariant conditions and " 132 "unswitch on them to eliminate branches that are " 133 "not-taken 1/<this option> times or less."), 134 cl::init(16)); 135 136 namespace { 137 struct CompareDesc { 138 BranchInst *Term; 139 Value *Invariant; 140 BasicBlock *InLoopSucc; 141 142 CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc) 143 : Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {} 144 }; 145 146 struct InjectedInvariant { 147 ICmpInst::Predicate Pred; 148 Value *LHS; 149 Value *RHS; 150 BasicBlock *InLoopSucc; 151 152 InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS, 153 BasicBlock *InLoopSucc) 154 : Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {} 155 }; 156 157 struct NonTrivialUnswitchCandidate { 158 Instruction *TI = nullptr; 159 TinyPtrVector<Value *> Invariants; 160 std::optional<InstructionCost> Cost; 161 std::optional<InjectedInvariant> PendingInjection; 162 NonTrivialUnswitchCandidate( 163 Instruction *TI, ArrayRef<Value *> Invariants, 164 std::optional<InstructionCost> Cost = std::nullopt, 165 std::optional<InjectedInvariant> PendingInjection = std::nullopt) 166 : TI(TI), Invariants(Invariants), Cost(Cost), 167 PendingInjection(PendingInjection) {}; 168 169 bool hasPendingInjection() const { return PendingInjection.has_value(); } 170 }; 171 } // end anonymous namespace. 172 173 // Helper to skip (select x, true, false), which matches both a logical AND and 174 // OR and can confuse code that tries to determine if \p Cond is either a 175 // logical AND or OR but not both. 176 static Value *skipTrivialSelect(Value *Cond) { 177 Value *CondNext; 178 while (match(Cond, m_Select(m_Value(CondNext), m_One(), m_Zero()))) 179 Cond = CondNext; 180 return Cond; 181 } 182 183 /// Collect all of the loop invariant input values transitively used by the 184 /// homogeneous instruction graph from a given root. 185 /// 186 /// This essentially walks from a root recursively through loop variant operands 187 /// which have perform the same logical operation (AND or OR) and finds all 188 /// inputs which are loop invariant. For some operations these can be 189 /// re-associated and unswitched out of the loop entirely. 190 static TinyPtrVector<Value *> 191 collectHomogenousInstGraphLoopInvariants(const Loop &L, Instruction &Root, 192 const LoopInfo &LI) { 193 assert(!L.isLoopInvariant(&Root) && 194 "Only need to walk the graph if root itself is not invariant."); 195 TinyPtrVector<Value *> Invariants; 196 197 bool IsRootAnd = match(&Root, m_LogicalAnd()); 198 bool IsRootOr = match(&Root, m_LogicalOr()); 199 200 // Build a worklist and recurse through operators collecting invariants. 201 SmallVector<Instruction *, 4> Worklist; 202 SmallPtrSet<Instruction *, 8> Visited; 203 Worklist.push_back(&Root); 204 Visited.insert(&Root); 205 do { 206 Instruction &I = *Worklist.pop_back_val(); 207 for (Value *OpV : I.operand_values()) { 208 // Skip constants as unswitching isn't interesting for them. 209 if (isa<Constant>(OpV)) 210 continue; 211 212 // Add it to our result if loop invariant. 213 if (L.isLoopInvariant(OpV)) { 214 Invariants.push_back(OpV); 215 continue; 216 } 217 218 // If not an instruction with the same opcode, nothing we can do. 219 Instruction *OpI = dyn_cast<Instruction>(skipTrivialSelect(OpV)); 220 221 if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) || 222 (IsRootOr && match(OpI, m_LogicalOr())))) { 223 // Visit this operand. 224 if (Visited.insert(OpI).second) 225 Worklist.push_back(OpI); 226 } 227 } 228 } while (!Worklist.empty()); 229 230 return Invariants; 231 } 232 233 static void replaceLoopInvariantUses(const Loop &L, Value *Invariant, 234 Constant &Replacement) { 235 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?"); 236 237 // Replace uses of LIC in the loop with the given constant. 238 // We use make_early_inc_range as set invalidates the iterator. 239 for (Use &U : llvm::make_early_inc_range(Invariant->uses())) { 240 Instruction *UserI = dyn_cast<Instruction>(U.getUser()); 241 242 // Replace this use within the loop body. 243 if (UserI && L.contains(UserI)) 244 U.set(&Replacement); 245 } 246 } 247 248 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial 249 /// incoming values along this edge. 250 static bool areLoopExitPHIsLoopInvariant(const Loop &L, 251 const BasicBlock &ExitingBB, 252 const BasicBlock &ExitBB) { 253 for (const Instruction &I : ExitBB) { 254 auto *PN = dyn_cast<PHINode>(&I); 255 if (!PN) 256 // No more PHIs to check. 257 return true; 258 259 // If the incoming value for this edge isn't loop invariant the unswitch 260 // won't be trivial. 261 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))) 262 return false; 263 } 264 llvm_unreachable("Basic blocks should never be empty!"); 265 } 266 267 /// Copy a set of loop invariant values \p ToDuplicate and insert them at the 268 /// end of \p BB and conditionally branch on the copied condition. We only 269 /// branch on a single value. 270 static void buildPartialUnswitchConditionalBranch( 271 BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction, 272 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze, 273 const Instruction *I, AssumptionCache *AC, const DominatorTree &DT) { 274 IRBuilder<> IRB(&BB); 275 276 SmallVector<Value *> FrozenInvariants; 277 for (Value *Inv : Invariants) { 278 if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(Inv, AC, I, &DT)) 279 Inv = IRB.CreateFreeze(Inv, Inv->getName() + ".fr"); 280 FrozenInvariants.push_back(Inv); 281 } 282 283 Value *Cond = Direction ? IRB.CreateOr(FrozenInvariants) 284 : IRB.CreateAnd(FrozenInvariants); 285 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, 286 Direction ? &NormalSucc : &UnswitchedSucc); 287 } 288 289 /// Copy a set of loop invariant values, and conditionally branch on them. 290 static void buildPartialInvariantUnswitchConditionalBranch( 291 BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction, 292 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L, 293 MemorySSAUpdater *MSSAU) { 294 ValueToValueMapTy VMap; 295 for (auto *Val : reverse(ToDuplicate)) { 296 Instruction *Inst = cast<Instruction>(Val); 297 Instruction *NewInst = Inst->clone(); 298 NewInst->insertInto(&BB, BB.end()); 299 RemapInstruction(NewInst, VMap, 300 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 301 VMap[Val] = NewInst; 302 303 if (!MSSAU) 304 continue; 305 306 MemorySSA *MSSA = MSSAU->getMemorySSA(); 307 if (auto *MemUse = 308 dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) { 309 auto *DefiningAccess = MemUse->getDefiningAccess(); 310 // Get the first defining access before the loop. 311 while (L.contains(DefiningAccess->getBlock())) { 312 // If the defining access is a MemoryPhi, get the incoming 313 // value for the pre-header as defining access. 314 if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess)) 315 DefiningAccess = 316 MemPhi->getIncomingValueForBlock(L.getLoopPreheader()); 317 else 318 DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess(); 319 } 320 MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess, 321 NewInst->getParent(), 322 MemorySSA::BeforeTerminator); 323 } 324 } 325 326 IRBuilder<> IRB(&BB); 327 Value *Cond = VMap[ToDuplicate[0]]; 328 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, 329 Direction ? &NormalSucc : &UnswitchedSucc); 330 } 331 332 /// Rewrite the PHI nodes in an unswitched loop exit basic block. 333 /// 334 /// Requires that the loop exit and unswitched basic block are the same, and 335 /// that the exiting block was a unique predecessor of that block. Rewrites the 336 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial 337 /// PHI nodes from the old preheader that now contains the unswitched 338 /// terminator. 339 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, 340 BasicBlock &OldExitingBB, 341 BasicBlock &OldPH) { 342 for (PHINode &PN : UnswitchedBB.phis()) { 343 // When the loop exit is directly unswitched we just need to update the 344 // incoming basic block. We loop to handle weird cases with repeated 345 // incoming blocks, but expect to typically only have one operand here. 346 for (auto i : seq<int>(0, PN.getNumOperands())) { 347 assert(PN.getIncomingBlock(i) == &OldExitingBB && 348 "Found incoming block different from unique predecessor!"); 349 PN.setIncomingBlock(i, &OldPH); 350 } 351 } 352 } 353 354 /// Rewrite the PHI nodes in the loop exit basic block and the split off 355 /// unswitched block. 356 /// 357 /// Because the exit block remains an exit from the loop, this rewrites the 358 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI 359 /// nodes into the unswitched basic block to select between the value in the 360 /// old preheader and the loop exit. 361 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, 362 BasicBlock &UnswitchedBB, 363 BasicBlock &OldExitingBB, 364 BasicBlock &OldPH, 365 bool FullUnswitch) { 366 assert(&ExitBB != &UnswitchedBB && 367 "Must have different loop exit and unswitched blocks!"); 368 BasicBlock::iterator InsertPt = UnswitchedBB.begin(); 369 for (PHINode &PN : ExitBB.phis()) { 370 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2, 371 PN.getName() + ".split"); 372 NewPN->insertBefore(InsertPt); 373 374 // Walk backwards over the old PHI node's inputs to minimize the cost of 375 // removing each one. We have to do this weird loop manually so that we 376 // create the same number of new incoming edges in the new PHI as we expect 377 // each case-based edge to be included in the unswitched switch in some 378 // cases. 379 // FIXME: This is really, really gross. It would be much cleaner if LLVM 380 // allowed us to create a single entry for a predecessor block without 381 // having separate entries for each "edge" even though these edges are 382 // required to produce identical results. 383 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) { 384 if (PN.getIncomingBlock(i) != &OldExitingBB) 385 continue; 386 387 Value *Incoming = PN.getIncomingValue(i); 388 if (FullUnswitch) 389 // No more edge from the old exiting block to the exit block. 390 PN.removeIncomingValue(i); 391 392 NewPN->addIncoming(Incoming, &OldPH); 393 } 394 395 // Now replace the old PHI with the new one and wire the old one in as an 396 // input to the new one. 397 PN.replaceAllUsesWith(NewPN); 398 NewPN->addIncoming(&PN, &ExitBB); 399 } 400 } 401 402 /// Hoist the current loop up to the innermost loop containing a remaining exit. 403 /// 404 /// Because we've removed an exit from the loop, we may have changed the set of 405 /// loops reachable and need to move the current loop up the loop nest or even 406 /// to an entirely separate nest. 407 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader, 408 DominatorTree &DT, LoopInfo &LI, 409 MemorySSAUpdater *MSSAU, ScalarEvolution *SE) { 410 // If the loop is already at the top level, we can't hoist it anywhere. 411 Loop *OldParentL = L.getParentLoop(); 412 if (!OldParentL) 413 return; 414 415 SmallVector<BasicBlock *, 4> Exits; 416 L.getExitBlocks(Exits); 417 Loop *NewParentL = nullptr; 418 for (auto *ExitBB : Exits) 419 if (Loop *ExitL = LI.getLoopFor(ExitBB)) 420 if (!NewParentL || NewParentL->contains(ExitL)) 421 NewParentL = ExitL; 422 423 if (NewParentL == OldParentL) 424 return; 425 426 // The new parent loop (if different) should always contain the old one. 427 if (NewParentL) 428 assert(NewParentL->contains(OldParentL) && 429 "Can only hoist this loop up the nest!"); 430 431 // The preheader will need to move with the body of this loop. However, 432 // because it isn't in this loop we also need to update the primary loop map. 433 assert(OldParentL == LI.getLoopFor(&Preheader) && 434 "Parent loop of this loop should contain this loop's preheader!"); 435 LI.changeLoopFor(&Preheader, NewParentL); 436 437 // Remove this loop from its old parent. 438 OldParentL->removeChildLoop(&L); 439 440 // Add the loop either to the new parent or as a top-level loop. 441 if (NewParentL) 442 NewParentL->addChildLoop(&L); 443 else 444 LI.addTopLevelLoop(&L); 445 446 // Remove this loops blocks from the old parent and every other loop up the 447 // nest until reaching the new parent. Also update all of these 448 // no-longer-containing loops to reflect the nesting change. 449 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL; 450 OldContainingL = OldContainingL->getParentLoop()) { 451 llvm::erase_if(OldContainingL->getBlocksVector(), 452 [&](const BasicBlock *BB) { 453 return BB == &Preheader || L.contains(BB); 454 }); 455 456 OldContainingL->getBlocksSet().erase(&Preheader); 457 for (BasicBlock *BB : L.blocks()) 458 OldContainingL->getBlocksSet().erase(BB); 459 460 // Because we just hoisted a loop out of this one, we have essentially 461 // created new exit paths from it. That means we need to form LCSSA PHI 462 // nodes for values used in the no-longer-nested loop. 463 formLCSSA(*OldContainingL, DT, &LI, SE); 464 465 // We shouldn't need to form dedicated exits because the exit introduced 466 // here is the (just split by unswitching) preheader. However, after trivial 467 // unswitching it is possible to get new non-dedicated exits out of parent 468 // loop so let's conservatively form dedicated exit blocks and figure out 469 // if we can optimize later. 470 formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU, 471 /*PreserveLCSSA*/ true); 472 } 473 } 474 475 // Return the top-most loop containing ExitBB and having ExitBB as exiting block 476 // or the loop containing ExitBB, if there is no parent loop containing ExitBB 477 // as exiting block. 478 static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB, 479 const LoopInfo &LI) { 480 Loop *TopMost = LI.getLoopFor(ExitBB); 481 Loop *Current = TopMost; 482 while (Current) { 483 if (Current->isLoopExiting(ExitBB)) 484 TopMost = Current; 485 Current = Current->getParentLoop(); 486 } 487 return TopMost; 488 } 489 490 /// Unswitch a trivial branch if the condition is loop invariant. 491 /// 492 /// This routine should only be called when loop code leading to the branch has 493 /// been validated as trivial (no side effects). This routine checks if the 494 /// condition is invariant and one of the successors is a loop exit. This 495 /// allows us to unswitch without duplicating the loop, making it trivial. 496 /// 497 /// If this routine fails to unswitch the branch it returns false. 498 /// 499 /// If the branch can be unswitched, this routine splits the preheader and 500 /// hoists the branch above that split. Preserves loop simplified form 501 /// (splitting the exit block as necessary). It simplifies the branch within 502 /// the loop to an unconditional branch but doesn't remove it entirely. Further 503 /// cleanup can be done with some simplifycfg like pass. 504 /// 505 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 506 /// invalidated by this. 507 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, 508 LoopInfo &LI, ScalarEvolution *SE, 509 MemorySSAUpdater *MSSAU) { 510 assert(BI.isConditional() && "Can only unswitch a conditional branch!"); 511 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n"); 512 513 // The loop invariant values that we want to unswitch. 514 TinyPtrVector<Value *> Invariants; 515 516 // When true, we're fully unswitching the branch rather than just unswitching 517 // some input conditions to the branch. 518 bool FullUnswitch = false; 519 520 Value *Cond = skipTrivialSelect(BI.getCondition()); 521 if (L.isLoopInvariant(Cond)) { 522 Invariants.push_back(Cond); 523 FullUnswitch = true; 524 } else { 525 if (auto *CondInst = dyn_cast<Instruction>(Cond)) 526 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI); 527 if (Invariants.empty()) { 528 LLVM_DEBUG(dbgs() << " Couldn't find invariant inputs!\n"); 529 return false; 530 } 531 } 532 533 // Check that one of the branch's successors exits, and which one. 534 bool ExitDirection = true; 535 int LoopExitSuccIdx = 0; 536 auto *LoopExitBB = BI.getSuccessor(0); 537 if (L.contains(LoopExitBB)) { 538 ExitDirection = false; 539 LoopExitSuccIdx = 1; 540 LoopExitBB = BI.getSuccessor(1); 541 if (L.contains(LoopExitBB)) { 542 LLVM_DEBUG(dbgs() << " Branch doesn't exit the loop!\n"); 543 return false; 544 } 545 } 546 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); 547 auto *ParentBB = BI.getParent(); 548 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) { 549 LLVM_DEBUG(dbgs() << " Loop exit PHI's aren't loop-invariant!\n"); 550 return false; 551 } 552 553 // When unswitching only part of the branch's condition, we need the exit 554 // block to be reached directly from the partially unswitched input. This can 555 // be done when the exit block is along the true edge and the branch condition 556 // is a graph of `or` operations, or the exit block is along the false edge 557 // and the condition is a graph of `and` operations. 558 if (!FullUnswitch) { 559 if (ExitDirection ? !match(Cond, m_LogicalOr()) 560 : !match(Cond, m_LogicalAnd())) { 561 LLVM_DEBUG(dbgs() << " Branch condition is in improper form for " 562 "non-full unswitch!\n"); 563 return false; 564 } 565 } 566 567 LLVM_DEBUG({ 568 dbgs() << " unswitching trivial invariant conditions for: " << BI 569 << "\n"; 570 for (Value *Invariant : Invariants) { 571 dbgs() << " " << *Invariant << " == true"; 572 if (Invariant != Invariants.back()) 573 dbgs() << " ||"; 574 dbgs() << "\n"; 575 } 576 }); 577 578 // If we have scalar evolutions, we need to invalidate them including this 579 // loop, the loop containing the exit block and the topmost parent loop 580 // exiting via LoopExitBB. 581 if (SE) { 582 if (const Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI)) 583 SE->forgetLoop(ExitL); 584 else 585 // Forget the entire nest as this exits the entire nest. 586 SE->forgetTopmostLoop(&L); 587 SE->forgetBlockAndLoopDispositions(); 588 } 589 590 if (MSSAU && VerifyMemorySSA) 591 MSSAU->getMemorySSA()->verifyMemorySSA(); 592 593 // Split the preheader, so that we know that there is a safe place to insert 594 // the conditional branch. We will change the preheader to have a conditional 595 // branch on LoopCond. 596 BasicBlock *OldPH = L.getLoopPreheader(); 597 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); 598 599 // Now that we have a place to insert the conditional branch, create a place 600 // to branch to: this is the exit block out of the loop that we are 601 // unswitching. We need to split this if there are other loop predecessors. 602 // Because the loop is in simplified form, *any* other predecessor is enough. 603 BasicBlock *UnswitchedBB; 604 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) { 605 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() && 606 "A branch's parent isn't a predecessor!"); 607 UnswitchedBB = LoopExitBB; 608 } else { 609 UnswitchedBB = 610 SplitBlock(LoopExitBB, LoopExitBB->begin(), &DT, &LI, MSSAU, "", false); 611 } 612 613 if (MSSAU && VerifyMemorySSA) 614 MSSAU->getMemorySSA()->verifyMemorySSA(); 615 616 // Actually move the invariant uses into the unswitched position. If possible, 617 // we do this by moving the instructions, but when doing partial unswitching 618 // we do it by building a new merge of the values in the unswitched position. 619 OldPH->getTerminator()->eraseFromParent(); 620 if (FullUnswitch) { 621 // If fully unswitching, we can use the existing branch instruction. 622 // Splice it into the old PH to gate reaching the new preheader and re-point 623 // its successors. 624 BI.moveBefore(*OldPH, OldPH->end()); 625 BI.setCondition(Cond); 626 if (MSSAU) { 627 // Temporarily clone the terminator, to make MSSA update cheaper by 628 // separating "insert edge" updates from "remove edge" ones. 629 BI.clone()->insertInto(ParentBB, ParentBB->end()); 630 } else { 631 // Create a new unconditional branch that will continue the loop as a new 632 // terminator. 633 BranchInst::Create(ContinueBB, ParentBB); 634 } 635 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); 636 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); 637 } else { 638 // Only unswitching a subset of inputs to the condition, so we will need to 639 // build a new branch that merges the invariant inputs. 640 if (ExitDirection) 641 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) && 642 "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the " 643 "condition!"); 644 else 645 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) && 646 "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the" 647 " condition!"); 648 buildPartialUnswitchConditionalBranch( 649 *OldPH, Invariants, ExitDirection, *UnswitchedBB, *NewPH, 650 FreezeLoopUnswitchCond, OldPH->getTerminator(), nullptr, DT); 651 } 652 653 // Update the dominator tree with the added edge. 654 DT.insertEdge(OldPH, UnswitchedBB); 655 656 // After the dominator tree was updated with the added edge, update MemorySSA 657 // if available. 658 if (MSSAU) { 659 SmallVector<CFGUpdate, 1> Updates; 660 Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB}); 661 MSSAU->applyInsertUpdates(Updates, DT); 662 } 663 664 // Finish updating dominator tree and memory ssa for full unswitch. 665 if (FullUnswitch) { 666 if (MSSAU) { 667 // Remove the cloned branch instruction. 668 ParentBB->getTerminator()->eraseFromParent(); 669 // Create unconditional branch now. 670 BranchInst::Create(ContinueBB, ParentBB); 671 MSSAU->removeEdge(ParentBB, LoopExitBB); 672 } 673 DT.deleteEdge(ParentBB, LoopExitBB); 674 } 675 676 if (MSSAU && VerifyMemorySSA) 677 MSSAU->getMemorySSA()->verifyMemorySSA(); 678 679 // Rewrite the relevant PHI nodes. 680 if (UnswitchedBB == LoopExitBB) 681 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); 682 else 683 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, 684 *ParentBB, *OldPH, FullUnswitch); 685 686 // The constant we can replace all of our invariants with inside the loop 687 // body. If any of the invariants have a value other than this the loop won't 688 // be entered. 689 ConstantInt *Replacement = ExitDirection 690 ? ConstantInt::getFalse(BI.getContext()) 691 : ConstantInt::getTrue(BI.getContext()); 692 693 // Since this is an i1 condition we can also trivially replace uses of it 694 // within the loop with a constant. 695 for (Value *Invariant : Invariants) 696 replaceLoopInvariantUses(L, Invariant, *Replacement); 697 698 // If this was full unswitching, we may have changed the nesting relationship 699 // for this loop so hoist it to its correct parent if needed. 700 if (FullUnswitch) 701 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE); 702 703 if (MSSAU && VerifyMemorySSA) 704 MSSAU->getMemorySSA()->verifyMemorySSA(); 705 706 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n"); 707 ++NumTrivial; 708 ++NumBranches; 709 return true; 710 } 711 712 /// Unswitch a trivial switch if the condition is loop invariant. 713 /// 714 /// This routine should only be called when loop code leading to the switch has 715 /// been validated as trivial (no side effects). This routine checks if the 716 /// condition is invariant and that at least one of the successors is a loop 717 /// exit. This allows us to unswitch without duplicating the loop, making it 718 /// trivial. 719 /// 720 /// If this routine fails to unswitch the switch it returns false. 721 /// 722 /// If the switch can be unswitched, this routine splits the preheader and 723 /// copies the switch above that split. If the default case is one of the 724 /// exiting cases, it copies the non-exiting cases and points them at the new 725 /// preheader. If the default case is not exiting, it copies the exiting cases 726 /// and points the default at the preheader. It preserves loop simplified form 727 /// (splitting the exit blocks as necessary). It simplifies the switch within 728 /// the loop by removing now-dead cases. If the default case is one of those 729 /// unswitched, it replaces its destination with a new basic block containing 730 /// only unreachable. Such basic blocks, while technically loop exits, are not 731 /// considered for unswitching so this is a stable transform and the same 732 /// switch will not be revisited. If after unswitching there is only a single 733 /// in-loop successor, the switch is further simplified to an unconditional 734 /// branch. Still more cleanup can be done with some simplifycfg like pass. 735 /// 736 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 737 /// invalidated by this. 738 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, 739 LoopInfo &LI, ScalarEvolution *SE, 740 MemorySSAUpdater *MSSAU) { 741 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n"); 742 Value *LoopCond = SI.getCondition(); 743 744 // If this isn't switching on an invariant condition, we can't unswitch it. 745 if (!L.isLoopInvariant(LoopCond)) 746 return false; 747 748 auto *ParentBB = SI.getParent(); 749 750 // The same check must be used both for the default and the exit cases. We 751 // should never leave edges from the switch instruction to a basic block that 752 // we are unswitching, hence the condition used to determine the default case 753 // needs to also be used to populate ExitCaseIndices, which is then used to 754 // remove cases from the switch. 755 auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) { 756 // BBToCheck is not an exit block if it is inside loop L. 757 if (L.contains(&BBToCheck)) 758 return false; 759 // BBToCheck is not trivial to unswitch if its phis aren't loop invariant. 760 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck)) 761 return false; 762 // We do not unswitch a block that only has an unreachable statement, as 763 // it's possible this is a previously unswitched block. Only unswitch if 764 // either the terminator is not unreachable, or, if it is, it's not the only 765 // instruction in the block. 766 auto *TI = BBToCheck.getTerminator(); 767 bool isUnreachable = isa<UnreachableInst>(TI); 768 return !isUnreachable || 769 (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI)); 770 }; 771 772 SmallVector<int, 4> ExitCaseIndices; 773 for (auto Case : SI.cases()) 774 if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor())) 775 ExitCaseIndices.push_back(Case.getCaseIndex()); 776 BasicBlock *DefaultExitBB = nullptr; 777 SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight = 778 SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0); 779 if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) { 780 DefaultExitBB = SI.getDefaultDest(); 781 } else if (ExitCaseIndices.empty()) 782 return false; 783 784 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n"); 785 786 if (MSSAU && VerifyMemorySSA) 787 MSSAU->getMemorySSA()->verifyMemorySSA(); 788 789 // We may need to invalidate SCEVs for the outermost loop reached by any of 790 // the exits. 791 Loop *OuterL = &L; 792 793 if (DefaultExitBB) { 794 // Check the loop containing this exit. 795 Loop *ExitL = getTopMostExitingLoop(DefaultExitBB, LI); 796 if (!ExitL || ExitL->contains(OuterL)) 797 OuterL = ExitL; 798 } 799 for (unsigned Index : ExitCaseIndices) { 800 auto CaseI = SI.case_begin() + Index; 801 // Compute the outer loop from this exit. 802 Loop *ExitL = getTopMostExitingLoop(CaseI->getCaseSuccessor(), LI); 803 if (!ExitL || ExitL->contains(OuterL)) 804 OuterL = ExitL; 805 } 806 807 if (SE) { 808 if (OuterL) 809 SE->forgetLoop(OuterL); 810 else 811 SE->forgetTopmostLoop(&L); 812 } 813 814 if (DefaultExitBB) { 815 // Clear out the default destination temporarily to allow accurate 816 // predecessor lists to be examined below. 817 SI.setDefaultDest(nullptr); 818 } 819 820 // Store the exit cases into a separate data structure and remove them from 821 // the switch. 822 SmallVector<std::tuple<ConstantInt *, BasicBlock *, 823 SwitchInstProfUpdateWrapper::CaseWeightOpt>, 824 4> ExitCases; 825 ExitCases.reserve(ExitCaseIndices.size()); 826 SwitchInstProfUpdateWrapper SIW(SI); 827 // We walk the case indices backwards so that we remove the last case first 828 // and don't disrupt the earlier indices. 829 for (unsigned Index : reverse(ExitCaseIndices)) { 830 auto CaseI = SI.case_begin() + Index; 831 // Save the value of this case. 832 auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex()); 833 ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W); 834 // Delete the unswitched cases. 835 SIW.removeCase(CaseI); 836 } 837 838 // Check if after this all of the remaining cases point at the same 839 // successor. 840 BasicBlock *CommonSuccBB = nullptr; 841 if (SI.getNumCases() > 0 && 842 all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) { 843 return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor(); 844 })) 845 CommonSuccBB = SI.case_begin()->getCaseSuccessor(); 846 if (!DefaultExitBB) { 847 // If we're not unswitching the default, we need it to match any cases to 848 // have a common successor or if we have no cases it is the common 849 // successor. 850 if (SI.getNumCases() == 0) 851 CommonSuccBB = SI.getDefaultDest(); 852 else if (SI.getDefaultDest() != CommonSuccBB) 853 CommonSuccBB = nullptr; 854 } 855 856 // Split the preheader, so that we know that there is a safe place to insert 857 // the switch. 858 BasicBlock *OldPH = L.getLoopPreheader(); 859 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); 860 OldPH->getTerminator()->eraseFromParent(); 861 862 // Now add the unswitched switch. 863 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); 864 SwitchInstProfUpdateWrapper NewSIW(*NewSI); 865 866 // Rewrite the IR for the unswitched basic blocks. This requires two steps. 867 // First, we split any exit blocks with remaining in-loop predecessors. Then 868 // we update the PHIs in one of two ways depending on if there was a split. 869 // We walk in reverse so that we split in the same order as the cases 870 // appeared. This is purely for convenience of reading the resulting IR, but 871 // it doesn't cost anything really. 872 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; 873 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; 874 // Handle the default exit if necessary. 875 // FIXME: It'd be great if we could merge this with the loop below but LLVM's 876 // ranges aren't quite powerful enough yet. 877 if (DefaultExitBB) { 878 if (pred_empty(DefaultExitBB)) { 879 UnswitchedExitBBs.insert(DefaultExitBB); 880 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); 881 } else { 882 auto *SplitBB = 883 SplitBlock(DefaultExitBB, DefaultExitBB->begin(), &DT, &LI, MSSAU); 884 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB, 885 *ParentBB, *OldPH, 886 /*FullUnswitch*/ true); 887 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; 888 } 889 } 890 // Note that we must use a reference in the for loop so that we update the 891 // container. 892 for (auto &ExitCase : reverse(ExitCases)) { 893 // Grab a reference to the exit block in the pair so that we can update it. 894 BasicBlock *ExitBB = std::get<1>(ExitCase); 895 896 // If this case is the last edge into the exit block, we can simply reuse it 897 // as it will no longer be a loop exit. No mapping necessary. 898 if (pred_empty(ExitBB)) { 899 // Only rewrite once. 900 if (UnswitchedExitBBs.insert(ExitBB).second) 901 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); 902 continue; 903 } 904 905 // Otherwise we need to split the exit block so that we retain an exit 906 // block from the loop and a target for the unswitched condition. 907 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; 908 if (!SplitExitBB) { 909 // If this is the first time we see this, do the split and remember it. 910 SplitExitBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU); 911 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB, 912 *ParentBB, *OldPH, 913 /*FullUnswitch*/ true); 914 } 915 // Update the case pair to point to the split block. 916 std::get<1>(ExitCase) = SplitExitBB; 917 } 918 919 // Now add the unswitched cases. We do this in reverse order as we built them 920 // in reverse order. 921 for (auto &ExitCase : reverse(ExitCases)) { 922 ConstantInt *CaseVal = std::get<0>(ExitCase); 923 BasicBlock *UnswitchedBB = std::get<1>(ExitCase); 924 925 NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase)); 926 } 927 928 // If the default was unswitched, re-point it and add explicit cases for 929 // entering the loop. 930 if (DefaultExitBB) { 931 NewSIW->setDefaultDest(DefaultExitBB); 932 NewSIW.setSuccessorWeight(0, DefaultCaseWeight); 933 934 // We removed all the exit cases, so we just copy the cases to the 935 // unswitched switch. 936 for (const auto &Case : SI.cases()) 937 NewSIW.addCase(Case.getCaseValue(), NewPH, 938 SIW.getSuccessorWeight(Case.getSuccessorIndex())); 939 } else if (DefaultCaseWeight) { 940 // We have to set branch weight of the default case. 941 uint64_t SW = *DefaultCaseWeight; 942 for (const auto &Case : SI.cases()) { 943 auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex()); 944 assert(W && 945 "case weight must be defined as default case weight is defined"); 946 SW += *W; 947 } 948 NewSIW.setSuccessorWeight(0, SW); 949 } 950 951 // If we ended up with a common successor for every path through the switch 952 // after unswitching, rewrite it to an unconditional branch to make it easy 953 // to recognize. Otherwise we potentially have to recognize the default case 954 // pointing at unreachable and other complexity. 955 if (CommonSuccBB) { 956 BasicBlock *BB = SI.getParent(); 957 // We may have had multiple edges to this common successor block, so remove 958 // them as predecessors. We skip the first one, either the default or the 959 // actual first case. 960 bool SkippedFirst = DefaultExitBB == nullptr; 961 for (auto Case : SI.cases()) { 962 assert(Case.getCaseSuccessor() == CommonSuccBB && 963 "Non-common successor!"); 964 (void)Case; 965 if (!SkippedFirst) { 966 SkippedFirst = true; 967 continue; 968 } 969 CommonSuccBB->removePredecessor(BB, 970 /*KeepOneInputPHIs*/ true); 971 } 972 // Now nuke the switch and replace it with a direct branch. 973 SIW.eraseFromParent(); 974 BranchInst::Create(CommonSuccBB, BB); 975 } else if (DefaultExitBB) { 976 assert(SI.getNumCases() > 0 && 977 "If we had no cases we'd have a common successor!"); 978 // Move the last case to the default successor. This is valid as if the 979 // default got unswitched it cannot be reached. This has the advantage of 980 // being simple and keeping the number of edges from this switch to 981 // successors the same, and avoiding any PHI update complexity. 982 auto LastCaseI = std::prev(SI.case_end()); 983 984 SI.setDefaultDest(LastCaseI->getCaseSuccessor()); 985 SIW.setSuccessorWeight( 986 0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex())); 987 SIW.removeCase(LastCaseI); 988 } 989 990 // Walk the unswitched exit blocks and the unswitched split blocks and update 991 // the dominator tree based on the CFG edits. While we are walking unordered 992 // containers here, the API for applyUpdates takes an unordered list of 993 // updates and requires them to not contain duplicates. 994 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 995 for (auto *UnswitchedExitBB : UnswitchedExitBBs) { 996 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB}); 997 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB}); 998 } 999 for (auto SplitUnswitchedPair : SplitExitBBMap) { 1000 DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first}); 1001 DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second}); 1002 } 1003 1004 if (MSSAU) { 1005 MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true); 1006 if (VerifyMemorySSA) 1007 MSSAU->getMemorySSA()->verifyMemorySSA(); 1008 } else { 1009 DT.applyUpdates(DTUpdates); 1010 } 1011 1012 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 1013 1014 // We may have changed the nesting relationship for this loop so hoist it to 1015 // its correct parent if needed. 1016 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE); 1017 1018 if (MSSAU && VerifyMemorySSA) 1019 MSSAU->getMemorySSA()->verifyMemorySSA(); 1020 1021 ++NumTrivial; 1022 ++NumSwitches; 1023 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n"); 1024 return true; 1025 } 1026 1027 /// This routine scans the loop to find a branch or switch which occurs before 1028 /// any side effects occur. These can potentially be unswitched without 1029 /// duplicating the loop. If a branch or switch is successfully unswitched the 1030 /// scanning continues to see if subsequent branches or switches have become 1031 /// trivial. Once all trivial candidates have been unswitched, this routine 1032 /// returns. 1033 /// 1034 /// The return value indicates whether anything was unswitched (and therefore 1035 /// changed). 1036 /// 1037 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 1038 /// invalidated by this. 1039 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, 1040 LoopInfo &LI, ScalarEvolution *SE, 1041 MemorySSAUpdater *MSSAU) { 1042 bool Changed = false; 1043 1044 // If loop header has only one reachable successor we should keep looking for 1045 // trivial condition candidates in the successor as well. An alternative is 1046 // to constant fold conditions and merge successors into loop header (then we 1047 // only need to check header's terminator). The reason for not doing this in 1048 // LoopUnswitch pass is that it could potentially break LoopPassManager's 1049 // invariants. Folding dead branches could either eliminate the current loop 1050 // or make other loops unreachable. LCSSA form might also not be preserved 1051 // after deleting branches. The following code keeps traversing loop header's 1052 // successors until it finds the trivial condition candidate (condition that 1053 // is not a constant). Since unswitching generates branches with constant 1054 // conditions, this scenario could be very common in practice. 1055 BasicBlock *CurrentBB = L.getHeader(); 1056 SmallPtrSet<BasicBlock *, 8> Visited; 1057 Visited.insert(CurrentBB); 1058 do { 1059 // Check if there are any side-effecting instructions (e.g. stores, calls, 1060 // volatile loads) in the part of the loop that the code *would* execute 1061 // without unswitching. 1062 if (MSSAU) // Possible early exit with MSSA 1063 if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB)) 1064 if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end())) 1065 return Changed; 1066 if (llvm::any_of(*CurrentBB, 1067 [](Instruction &I) { return I.mayHaveSideEffects(); })) 1068 return Changed; 1069 1070 Instruction *CurrentTerm = CurrentBB->getTerminator(); 1071 1072 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) { 1073 // Don't bother trying to unswitch past a switch with a constant 1074 // condition. This should be removed prior to running this pass by 1075 // simplifycfg. 1076 if (isa<Constant>(SI->getCondition())) 1077 return Changed; 1078 1079 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU)) 1080 // Couldn't unswitch this one so we're done. 1081 return Changed; 1082 1083 // Mark that we managed to unswitch something. 1084 Changed = true; 1085 1086 // If unswitching turned the terminator into an unconditional branch then 1087 // we can continue. The unswitching logic specifically works to fold any 1088 // cases it can into an unconditional branch to make it easier to 1089 // recognize here. 1090 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); 1091 if (!BI || BI->isConditional()) 1092 return Changed; 1093 1094 CurrentBB = BI->getSuccessor(0); 1095 continue; 1096 } 1097 1098 auto *BI = dyn_cast<BranchInst>(CurrentTerm); 1099 if (!BI) 1100 // We do not understand other terminator instructions. 1101 return Changed; 1102 1103 // Don't bother trying to unswitch past an unconditional branch or a branch 1104 // with a constant value. These should be removed by simplifycfg prior to 1105 // running this pass. 1106 if (!BI->isConditional() || 1107 isa<Constant>(skipTrivialSelect(BI->getCondition()))) 1108 return Changed; 1109 1110 // Found a trivial condition candidate: non-foldable conditional branch. If 1111 // we fail to unswitch this, we can't do anything else that is trivial. 1112 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU)) 1113 return Changed; 1114 1115 // Mark that we managed to unswitch something. 1116 Changed = true; 1117 1118 // If we only unswitched some of the conditions feeding the branch, we won't 1119 // have collapsed it to a single successor. 1120 BI = cast<BranchInst>(CurrentBB->getTerminator()); 1121 if (BI->isConditional()) 1122 return Changed; 1123 1124 // Follow the newly unconditional branch into its successor. 1125 CurrentBB = BI->getSuccessor(0); 1126 1127 // When continuing, if we exit the loop or reach a previous visited block, 1128 // then we can not reach any trivial condition candidates (unfoldable 1129 // branch instructions or switch instructions) and no unswitch can happen. 1130 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second); 1131 1132 return Changed; 1133 } 1134 1135 /// Build the cloned blocks for an unswitched copy of the given loop. 1136 /// 1137 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and 1138 /// after the split block (`SplitBB`) that will be used to select between the 1139 /// cloned and original loop. 1140 /// 1141 /// This routine handles cloning all of the necessary loop blocks and exit 1142 /// blocks including rewriting their instructions and the relevant PHI nodes. 1143 /// Any loop blocks or exit blocks which are dominated by a different successor 1144 /// than the one for this clone of the loop blocks can be trivially skipped. We 1145 /// use the `DominatingSucc` map to determine whether a block satisfies that 1146 /// property with a simple map lookup. 1147 /// 1148 /// It also correctly creates the unconditional branch in the cloned 1149 /// unswitched parent block to only point at the unswitched successor. 1150 /// 1151 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit 1152 /// block splitting is correctly reflected in `LoopInfo`, essentially all of 1153 /// the cloned blocks (and their loops) are left without full `LoopInfo` 1154 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned 1155 /// blocks to them but doesn't create the cloned `DominatorTree` structure and 1156 /// instead the caller must recompute an accurate DT. It *does* correctly 1157 /// update the `AssumptionCache` provided in `AC`. 1158 static BasicBlock *buildClonedLoopBlocks( 1159 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, 1160 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB, 1161 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, 1162 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc, 1163 ValueToValueMapTy &VMap, 1164 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC, 1165 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU, 1166 ScalarEvolution *SE) { 1167 SmallVector<BasicBlock *, 4> NewBlocks; 1168 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size()); 1169 1170 // We will need to clone a bunch of blocks, wrap up the clone operation in 1171 // a helper. 1172 auto CloneBlock = [&](BasicBlock *OldBB) { 1173 // Clone the basic block and insert it before the new preheader. 1174 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent()); 1175 NewBB->moveBefore(LoopPH); 1176 1177 // Record this block and the mapping. 1178 NewBlocks.push_back(NewBB); 1179 VMap[OldBB] = NewBB; 1180 1181 return NewBB; 1182 }; 1183 1184 // We skip cloning blocks when they have a dominating succ that is not the 1185 // succ we are cloning for. 1186 auto SkipBlock = [&](BasicBlock *BB) { 1187 auto It = DominatingSucc.find(BB); 1188 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB; 1189 }; 1190 1191 // First, clone the preheader. 1192 auto *ClonedPH = CloneBlock(LoopPH); 1193 1194 // Then clone all the loop blocks, skipping the ones that aren't necessary. 1195 for (auto *LoopBB : L.blocks()) 1196 if (!SkipBlock(LoopBB)) 1197 CloneBlock(LoopBB); 1198 1199 // Split all the loop exit edges so that when we clone the exit blocks, if 1200 // any of the exit blocks are *also* a preheader for some other loop, we 1201 // don't create multiple predecessors entering the loop header. 1202 for (auto *ExitBB : ExitBlocks) { 1203 if (SkipBlock(ExitBB)) 1204 continue; 1205 1206 // When we are going to clone an exit, we don't need to clone all the 1207 // instructions in the exit block and we want to ensure we have an easy 1208 // place to merge the CFG, so split the exit first. This is always safe to 1209 // do because there cannot be any non-loop predecessors of a loop exit in 1210 // loop simplified form. 1211 auto *MergeBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU); 1212 1213 // Rearrange the names to make it easier to write test cases by having the 1214 // exit block carry the suffix rather than the merge block carrying the 1215 // suffix. 1216 MergeBB->takeName(ExitBB); 1217 ExitBB->setName(Twine(MergeBB->getName()) + ".split"); 1218 1219 // Now clone the original exit block. 1220 auto *ClonedExitBB = CloneBlock(ExitBB); 1221 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && 1222 "Exit block should have been split to have one successor!"); 1223 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && 1224 "Cloned exit block has the wrong successor!"); 1225 1226 // Remap any cloned instructions and create a merge phi node for them. 1227 for (auto ZippedInsts : llvm::zip_first( 1228 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())), 1229 llvm::make_range(ClonedExitBB->begin(), 1230 std::prev(ClonedExitBB->end())))) { 1231 Instruction &I = std::get<0>(ZippedInsts); 1232 Instruction &ClonedI = std::get<1>(ZippedInsts); 1233 1234 // The only instructions in the exit block should be PHI nodes and 1235 // potentially a landing pad. 1236 assert( 1237 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && 1238 "Bad instruction in exit block!"); 1239 // We should have a value map between the instruction and its clone. 1240 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!"); 1241 1242 // Forget SCEVs based on exit phis in case SCEV looked through the phi. 1243 if (SE && isa<PHINode>(I)) 1244 SE->forgetValue(&I); 1245 1246 auto *MergePN = 1247 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi"); 1248 MergePN->insertBefore(MergeBB->getFirstInsertionPt()); 1249 I.replaceAllUsesWith(MergePN); 1250 MergePN->addIncoming(&I, ExitBB); 1251 MergePN->addIncoming(&ClonedI, ClonedExitBB); 1252 } 1253 } 1254 1255 // Rewrite the instructions in the cloned blocks to refer to the instructions 1256 // in the cloned blocks. We have to do this as a second pass so that we have 1257 // everything available. Also, we have inserted new instructions which may 1258 // include assume intrinsics, so we update the assumption cache while 1259 // processing this. 1260 Module *M = ClonedPH->getParent()->getParent(); 1261 for (auto *ClonedBB : NewBlocks) 1262 for (Instruction &I : *ClonedBB) { 1263 RemapDPValueRange(M, I.getDbgValueRange(), VMap, 1264 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1265 RemapInstruction(&I, VMap, 1266 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1267 if (auto *II = dyn_cast<AssumeInst>(&I)) 1268 AC.registerAssumption(II); 1269 } 1270 1271 // Update any PHI nodes in the cloned successors of the skipped blocks to not 1272 // have spurious incoming values. 1273 for (auto *LoopBB : L.blocks()) 1274 if (SkipBlock(LoopBB)) 1275 for (auto *SuccBB : successors(LoopBB)) 1276 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB))) 1277 for (PHINode &PN : ClonedSuccBB->phis()) 1278 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false); 1279 1280 // Remove the cloned parent as a predecessor of any successor we ended up 1281 // cloning other than the unswitched one. 1282 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB)); 1283 for (auto *SuccBB : successors(ParentBB)) { 1284 if (SuccBB == UnswitchedSuccBB) 1285 continue; 1286 1287 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)); 1288 if (!ClonedSuccBB) 1289 continue; 1290 1291 ClonedSuccBB->removePredecessor(ClonedParentBB, 1292 /*KeepOneInputPHIs*/ true); 1293 } 1294 1295 // Replace the cloned branch with an unconditional branch to the cloned 1296 // unswitched successor. 1297 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB)); 1298 Instruction *ClonedTerminator = ClonedParentBB->getTerminator(); 1299 // Trivial Simplification. If Terminator is a conditional branch and 1300 // condition becomes dead - erase it. 1301 Value *ClonedConditionToErase = nullptr; 1302 if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator)) 1303 ClonedConditionToErase = BI->getCondition(); 1304 else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator)) 1305 ClonedConditionToErase = SI->getCondition(); 1306 1307 ClonedTerminator->eraseFromParent(); 1308 BranchInst::Create(ClonedSuccBB, ClonedParentBB); 1309 1310 if (ClonedConditionToErase) 1311 RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr, 1312 MSSAU); 1313 1314 // If there are duplicate entries in the PHI nodes because of multiple edges 1315 // to the unswitched successor, we need to nuke all but one as we replaced it 1316 // with a direct branch. 1317 for (PHINode &PN : ClonedSuccBB->phis()) { 1318 bool Found = false; 1319 // Loop over the incoming operands backwards so we can easily delete as we 1320 // go without invalidating the index. 1321 for (int i = PN.getNumOperands() - 1; i >= 0; --i) { 1322 if (PN.getIncomingBlock(i) != ClonedParentBB) 1323 continue; 1324 if (!Found) { 1325 Found = true; 1326 continue; 1327 } 1328 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false); 1329 } 1330 } 1331 1332 // Record the domtree updates for the new blocks. 1333 SmallPtrSet<BasicBlock *, 4> SuccSet; 1334 for (auto *ClonedBB : NewBlocks) { 1335 for (auto *SuccBB : successors(ClonedBB)) 1336 if (SuccSet.insert(SuccBB).second) 1337 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB}); 1338 SuccSet.clear(); 1339 } 1340 1341 return ClonedPH; 1342 } 1343 1344 /// Recursively clone the specified loop and all of its children. 1345 /// 1346 /// The target parent loop for the clone should be provided, or can be null if 1347 /// the clone is a top-level loop. While cloning, all the blocks are mapped 1348 /// with the provided value map. The entire original loop must be present in 1349 /// the value map. The cloned loop is returned. 1350 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, 1351 const ValueToValueMapTy &VMap, LoopInfo &LI) { 1352 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) { 1353 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!"); 1354 ClonedL.reserveBlocks(OrigL.getNumBlocks()); 1355 for (auto *BB : OrigL.blocks()) { 1356 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB)); 1357 ClonedL.addBlockEntry(ClonedBB); 1358 if (LI.getLoopFor(BB) == &OrigL) 1359 LI.changeLoopFor(ClonedBB, &ClonedL); 1360 } 1361 }; 1362 1363 // We specially handle the first loop because it may get cloned into 1364 // a different parent and because we most commonly are cloning leaf loops. 1365 Loop *ClonedRootL = LI.AllocateLoop(); 1366 if (RootParentL) 1367 RootParentL->addChildLoop(ClonedRootL); 1368 else 1369 LI.addTopLevelLoop(ClonedRootL); 1370 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL); 1371 1372 if (OrigRootL.isInnermost()) 1373 return ClonedRootL; 1374 1375 // If we have a nest, we can quickly clone the entire loop nest using an 1376 // iterative approach because it is a tree. We keep the cloned parent in the 1377 // data structure to avoid repeatedly querying through a map to find it. 1378 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone; 1379 // Build up the loops to clone in reverse order as we'll clone them from the 1380 // back. 1381 for (Loop *ChildL : llvm::reverse(OrigRootL)) 1382 LoopsToClone.push_back({ClonedRootL, ChildL}); 1383 do { 1384 Loop *ClonedParentL, *L; 1385 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val(); 1386 Loop *ClonedL = LI.AllocateLoop(); 1387 ClonedParentL->addChildLoop(ClonedL); 1388 AddClonedBlocksToLoop(*L, *ClonedL); 1389 for (Loop *ChildL : llvm::reverse(*L)) 1390 LoopsToClone.push_back({ClonedL, ChildL}); 1391 } while (!LoopsToClone.empty()); 1392 1393 return ClonedRootL; 1394 } 1395 1396 /// Build the cloned loops of an original loop from unswitching. 1397 /// 1398 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial 1399 /// operation. We need to re-verify that there even is a loop (as the backedge 1400 /// may not have been cloned), and even if there are remaining backedges the 1401 /// backedge set may be different. However, we know that each child loop is 1402 /// undisturbed, we only need to find where to place each child loop within 1403 /// either any parent loop or within a cloned version of the original loop. 1404 /// 1405 /// Because child loops may end up cloned outside of any cloned version of the 1406 /// original loop, multiple cloned sibling loops may be created. All of them 1407 /// are returned so that the newly introduced loop nest roots can be 1408 /// identified. 1409 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks, 1410 const ValueToValueMapTy &VMap, LoopInfo &LI, 1411 SmallVectorImpl<Loop *> &NonChildClonedLoops) { 1412 Loop *ClonedL = nullptr; 1413 1414 auto *OrigPH = OrigL.getLoopPreheader(); 1415 auto *OrigHeader = OrigL.getHeader(); 1416 1417 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH)); 1418 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader)); 1419 1420 // We need to know the loops of the cloned exit blocks to even compute the 1421 // accurate parent loop. If we only clone exits to some parent of the 1422 // original parent, we want to clone into that outer loop. We also keep track 1423 // of the loops that our cloned exit blocks participate in. 1424 Loop *ParentL = nullptr; 1425 SmallVector<BasicBlock *, 4> ClonedExitsInLoops; 1426 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap; 1427 ClonedExitsInLoops.reserve(ExitBlocks.size()); 1428 for (auto *ExitBB : ExitBlocks) 1429 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB))) 1430 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1431 ExitLoopMap[ClonedExitBB] = ExitL; 1432 ClonedExitsInLoops.push_back(ClonedExitBB); 1433 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1434 ParentL = ExitL; 1435 } 1436 assert((!ParentL || ParentL == OrigL.getParentLoop() || 1437 ParentL->contains(OrigL.getParentLoop())) && 1438 "The computed parent loop should always contain (or be) the parent of " 1439 "the original loop."); 1440 1441 // We build the set of blocks dominated by the cloned header from the set of 1442 // cloned blocks out of the original loop. While not all of these will 1443 // necessarily be in the cloned loop, it is enough to establish that they 1444 // aren't in unreachable cycles, etc. 1445 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks; 1446 for (auto *BB : OrigL.blocks()) 1447 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB))) 1448 ClonedLoopBlocks.insert(ClonedBB); 1449 1450 // Rebuild the set of blocks that will end up in the cloned loop. We may have 1451 // skipped cloning some region of this loop which can in turn skip some of 1452 // the backedges so we have to rebuild the blocks in the loop based on the 1453 // backedges that remain after cloning. 1454 SmallVector<BasicBlock *, 16> Worklist; 1455 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop; 1456 for (auto *Pred : predecessors(ClonedHeader)) { 1457 // The only possible non-loop header predecessor is the preheader because 1458 // we know we cloned the loop in simplified form. 1459 if (Pred == ClonedPH) 1460 continue; 1461 1462 // Because the loop was in simplified form, the only non-loop predecessor 1463 // should be the preheader. 1464 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop " 1465 "header other than the preheader " 1466 "that is not part of the loop!"); 1467 1468 // Insert this block into the loop set and on the first visit (and if it 1469 // isn't the header we're currently walking) put it into the worklist to 1470 // recurse through. 1471 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader) 1472 Worklist.push_back(Pred); 1473 } 1474 1475 // If we had any backedges then there *is* a cloned loop. Put the header into 1476 // the loop set and then walk the worklist backwards to find all the blocks 1477 // that remain within the loop after cloning. 1478 if (!BlocksInClonedLoop.empty()) { 1479 BlocksInClonedLoop.insert(ClonedHeader); 1480 1481 while (!Worklist.empty()) { 1482 BasicBlock *BB = Worklist.pop_back_val(); 1483 assert(BlocksInClonedLoop.count(BB) && 1484 "Didn't put block into the loop set!"); 1485 1486 // Insert any predecessors that are in the possible set into the cloned 1487 // set, and if the insert is successful, add them to the worklist. Note 1488 // that we filter on the blocks that are definitely reachable via the 1489 // backedge to the loop header so we may prune out dead code within the 1490 // cloned loop. 1491 for (auto *Pred : predecessors(BB)) 1492 if (ClonedLoopBlocks.count(Pred) && 1493 BlocksInClonedLoop.insert(Pred).second) 1494 Worklist.push_back(Pred); 1495 } 1496 1497 ClonedL = LI.AllocateLoop(); 1498 if (ParentL) { 1499 ParentL->addBasicBlockToLoop(ClonedPH, LI); 1500 ParentL->addChildLoop(ClonedL); 1501 } else { 1502 LI.addTopLevelLoop(ClonedL); 1503 } 1504 NonChildClonedLoops.push_back(ClonedL); 1505 1506 ClonedL->reserveBlocks(BlocksInClonedLoop.size()); 1507 // We don't want to just add the cloned loop blocks based on how we 1508 // discovered them. The original order of blocks was carefully built in 1509 // a way that doesn't rely on predecessor ordering. Rather than re-invent 1510 // that logic, we just re-walk the original blocks (and those of the child 1511 // loops) and filter them as we add them into the cloned loop. 1512 for (auto *BB : OrigL.blocks()) { 1513 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)); 1514 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB)) 1515 continue; 1516 1517 // Directly add the blocks that are only in this loop. 1518 if (LI.getLoopFor(BB) == &OrigL) { 1519 ClonedL->addBasicBlockToLoop(ClonedBB, LI); 1520 continue; 1521 } 1522 1523 // We want to manually add it to this loop and parents. 1524 // Registering it with LoopInfo will happen when we clone the top 1525 // loop for this block. 1526 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop()) 1527 PL->addBlockEntry(ClonedBB); 1528 } 1529 1530 // Now add each child loop whose header remains within the cloned loop. All 1531 // of the blocks within the loop must satisfy the same constraints as the 1532 // header so once we pass the header checks we can just clone the entire 1533 // child loop nest. 1534 for (Loop *ChildL : OrigL) { 1535 auto *ClonedChildHeader = 1536 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1537 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader)) 1538 continue; 1539 1540 #ifndef NDEBUG 1541 // We should never have a cloned child loop header but fail to have 1542 // all of the blocks for that child loop. 1543 for (auto *ChildLoopBB : ChildL->blocks()) 1544 assert(BlocksInClonedLoop.count( 1545 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && 1546 "Child cloned loop has a header within the cloned outer " 1547 "loop but not all of its blocks!"); 1548 #endif 1549 1550 cloneLoopNest(*ChildL, ClonedL, VMap, LI); 1551 } 1552 } 1553 1554 // Now that we've handled all the components of the original loop that were 1555 // cloned into a new loop, we still need to handle anything from the original 1556 // loop that wasn't in a cloned loop. 1557 1558 // Figure out what blocks are left to place within any loop nest containing 1559 // the unswitched loop. If we never formed a loop, the cloned PH is one of 1560 // them. 1561 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet; 1562 if (BlocksInClonedLoop.empty()) 1563 UnloopedBlockSet.insert(ClonedPH); 1564 for (auto *ClonedBB : ClonedLoopBlocks) 1565 if (!BlocksInClonedLoop.count(ClonedBB)) 1566 UnloopedBlockSet.insert(ClonedBB); 1567 1568 // Copy the cloned exits and sort them in ascending loop depth, we'll work 1569 // backwards across these to process them inside out. The order shouldn't 1570 // matter as we're just trying to build up the map from inside-out; we use 1571 // the map in a more stably ordered way below. 1572 auto OrderedClonedExitsInLoops = ClonedExitsInLoops; 1573 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { 1574 return ExitLoopMap.lookup(LHS)->getLoopDepth() < 1575 ExitLoopMap.lookup(RHS)->getLoopDepth(); 1576 }); 1577 1578 // Populate the existing ExitLoopMap with everything reachable from each 1579 // exit, starting from the inner most exit. 1580 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) { 1581 assert(Worklist.empty() && "Didn't clear worklist!"); 1582 1583 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val(); 1584 Loop *ExitL = ExitLoopMap.lookup(ExitBB); 1585 1586 // Walk the CFG back until we hit the cloned PH adding everything reachable 1587 // and in the unlooped set to this exit block's loop. 1588 Worklist.push_back(ExitBB); 1589 do { 1590 BasicBlock *BB = Worklist.pop_back_val(); 1591 // We can stop recursing at the cloned preheader (if we get there). 1592 if (BB == ClonedPH) 1593 continue; 1594 1595 for (BasicBlock *PredBB : predecessors(BB)) { 1596 // If this pred has already been moved to our set or is part of some 1597 // (inner) loop, no update needed. 1598 if (!UnloopedBlockSet.erase(PredBB)) { 1599 assert( 1600 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && 1601 "Predecessor not mapped to a loop!"); 1602 continue; 1603 } 1604 1605 // We just insert into the loop set here. We'll add these blocks to the 1606 // exit loop after we build up the set in an order that doesn't rely on 1607 // predecessor order (which in turn relies on use list order). 1608 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second; 1609 (void)Inserted; 1610 assert(Inserted && "Should only visit an unlooped block once!"); 1611 1612 // And recurse through to its predecessors. 1613 Worklist.push_back(PredBB); 1614 } 1615 } while (!Worklist.empty()); 1616 } 1617 1618 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned 1619 // blocks to their outer loops, walk the cloned blocks and the cloned exits 1620 // in their original order adding them to the correct loop. 1621 1622 // We need a stable insertion order. We use the order of the original loop 1623 // order and map into the correct parent loop. 1624 for (auto *BB : llvm::concat<BasicBlock *const>( 1625 ArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops)) 1626 if (Loop *OuterL = ExitLoopMap.lookup(BB)) 1627 OuterL->addBasicBlockToLoop(BB, LI); 1628 1629 #ifndef NDEBUG 1630 for (auto &BBAndL : ExitLoopMap) { 1631 auto *BB = BBAndL.first; 1632 auto *OuterL = BBAndL.second; 1633 assert(LI.getLoopFor(BB) == OuterL && 1634 "Failed to put all blocks into outer loops!"); 1635 } 1636 #endif 1637 1638 // Now that all the blocks are placed into the correct containing loop in the 1639 // absence of child loops, find all the potentially cloned child loops and 1640 // clone them into whatever outer loop we placed their header into. 1641 for (Loop *ChildL : OrigL) { 1642 auto *ClonedChildHeader = 1643 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1644 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader)) 1645 continue; 1646 1647 #ifndef NDEBUG 1648 for (auto *ChildLoopBB : ChildL->blocks()) 1649 assert(VMap.count(ChildLoopBB) && 1650 "Cloned a child loop header but not all of that loops blocks!"); 1651 #endif 1652 1653 NonChildClonedLoops.push_back(cloneLoopNest( 1654 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI)); 1655 } 1656 } 1657 1658 static void 1659 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1660 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, 1661 DominatorTree &DT, MemorySSAUpdater *MSSAU) { 1662 // Find all the dead clones, and remove them from their successors. 1663 SmallVector<BasicBlock *, 16> DeadBlocks; 1664 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks)) 1665 for (const auto &VMap : VMaps) 1666 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB))) 1667 if (!DT.isReachableFromEntry(ClonedBB)) { 1668 for (BasicBlock *SuccBB : successors(ClonedBB)) 1669 SuccBB->removePredecessor(ClonedBB); 1670 DeadBlocks.push_back(ClonedBB); 1671 } 1672 1673 // Remove all MemorySSA in the dead blocks 1674 if (MSSAU) { 1675 SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(), 1676 DeadBlocks.end()); 1677 MSSAU->removeBlocks(DeadBlockSet); 1678 } 1679 1680 // Drop any remaining references to break cycles. 1681 for (BasicBlock *BB : DeadBlocks) 1682 BB->dropAllReferences(); 1683 // Erase them from the IR. 1684 for (BasicBlock *BB : DeadBlocks) 1685 BB->eraseFromParent(); 1686 } 1687 1688 static void deleteDeadBlocksFromLoop(Loop &L, 1689 SmallVectorImpl<BasicBlock *> &ExitBlocks, 1690 DominatorTree &DT, LoopInfo &LI, 1691 MemorySSAUpdater *MSSAU, 1692 ScalarEvolution *SE, 1693 LPMUpdater &LoopUpdater) { 1694 // Find all the dead blocks tied to this loop, and remove them from their 1695 // successors. 1696 SmallSetVector<BasicBlock *, 8> DeadBlockSet; 1697 1698 // Start with loop/exit blocks and get a transitive closure of reachable dead 1699 // blocks. 1700 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(), 1701 ExitBlocks.end()); 1702 DeathCandidates.append(L.blocks().begin(), L.blocks().end()); 1703 while (!DeathCandidates.empty()) { 1704 auto *BB = DeathCandidates.pop_back_val(); 1705 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) { 1706 for (BasicBlock *SuccBB : successors(BB)) { 1707 SuccBB->removePredecessor(BB); 1708 DeathCandidates.push_back(SuccBB); 1709 } 1710 DeadBlockSet.insert(BB); 1711 } 1712 } 1713 1714 // Remove all MemorySSA in the dead blocks 1715 if (MSSAU) 1716 MSSAU->removeBlocks(DeadBlockSet); 1717 1718 // Filter out the dead blocks from the exit blocks list so that it can be 1719 // used in the caller. 1720 llvm::erase_if(ExitBlocks, 1721 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1722 1723 // Walk from this loop up through its parents removing all of the dead blocks. 1724 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) { 1725 for (auto *BB : DeadBlockSet) 1726 ParentL->getBlocksSet().erase(BB); 1727 llvm::erase_if(ParentL->getBlocksVector(), 1728 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1729 } 1730 1731 // Now delete the dead child loops. This raw delete will clear them 1732 // recursively. 1733 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) { 1734 if (!DeadBlockSet.count(ChildL->getHeader())) 1735 return false; 1736 1737 assert(llvm::all_of(ChildL->blocks(), 1738 [&](BasicBlock *ChildBB) { 1739 return DeadBlockSet.count(ChildBB); 1740 }) && 1741 "If the child loop header is dead all blocks in the child loop must " 1742 "be dead as well!"); 1743 LoopUpdater.markLoopAsDeleted(*ChildL, ChildL->getName()); 1744 if (SE) 1745 SE->forgetBlockAndLoopDispositions(); 1746 LI.destroy(ChildL); 1747 return true; 1748 }); 1749 1750 // Remove the loop mappings for the dead blocks and drop all the references 1751 // from these blocks to others to handle cyclic references as we start 1752 // deleting the blocks themselves. 1753 for (auto *BB : DeadBlockSet) { 1754 // Check that the dominator tree has already been updated. 1755 assert(!DT.getNode(BB) && "Should already have cleared domtree!"); 1756 LI.changeLoopFor(BB, nullptr); 1757 // Drop all uses of the instructions to make sure we won't have dangling 1758 // uses in other blocks. 1759 for (auto &I : *BB) 1760 if (!I.use_empty()) 1761 I.replaceAllUsesWith(PoisonValue::get(I.getType())); 1762 BB->dropAllReferences(); 1763 } 1764 1765 // Actually delete the blocks now that they've been fully unhooked from the 1766 // IR. 1767 for (auto *BB : DeadBlockSet) 1768 BB->eraseFromParent(); 1769 } 1770 1771 /// Recompute the set of blocks in a loop after unswitching. 1772 /// 1773 /// This walks from the original headers predecessors to rebuild the loop. We 1774 /// take advantage of the fact that new blocks can't have been added, and so we 1775 /// filter by the original loop's blocks. This also handles potentially 1776 /// unreachable code that we don't want to explore but might be found examining 1777 /// the predecessors of the header. 1778 /// 1779 /// If the original loop is no longer a loop, this will return an empty set. If 1780 /// it remains a loop, all the blocks within it will be added to the set 1781 /// (including those blocks in inner loops). 1782 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L, 1783 LoopInfo &LI) { 1784 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet; 1785 1786 auto *PH = L.getLoopPreheader(); 1787 auto *Header = L.getHeader(); 1788 1789 // A worklist to use while walking backwards from the header. 1790 SmallVector<BasicBlock *, 16> Worklist; 1791 1792 // First walk the predecessors of the header to find the backedges. This will 1793 // form the basis of our walk. 1794 for (auto *Pred : predecessors(Header)) { 1795 // Skip the preheader. 1796 if (Pred == PH) 1797 continue; 1798 1799 // Because the loop was in simplified form, the only non-loop predecessor 1800 // is the preheader. 1801 assert(L.contains(Pred) && "Found a predecessor of the loop header other " 1802 "than the preheader that is not part of the " 1803 "loop!"); 1804 1805 // Insert this block into the loop set and on the first visit and, if it 1806 // isn't the header we're currently walking, put it into the worklist to 1807 // recurse through. 1808 if (LoopBlockSet.insert(Pred).second && Pred != Header) 1809 Worklist.push_back(Pred); 1810 } 1811 1812 // If no backedges were found, we're done. 1813 if (LoopBlockSet.empty()) 1814 return LoopBlockSet; 1815 1816 // We found backedges, recurse through them to identify the loop blocks. 1817 while (!Worklist.empty()) { 1818 BasicBlock *BB = Worklist.pop_back_val(); 1819 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!"); 1820 1821 // No need to walk past the header. 1822 if (BB == Header) 1823 continue; 1824 1825 // Because we know the inner loop structure remains valid we can use the 1826 // loop structure to jump immediately across the entire nested loop. 1827 // Further, because it is in loop simplified form, we can directly jump 1828 // to its preheader afterward. 1829 if (Loop *InnerL = LI.getLoopFor(BB)) 1830 if (InnerL != &L) { 1831 assert(L.contains(InnerL) && 1832 "Should not reach a loop *outside* this loop!"); 1833 // The preheader is the only possible predecessor of the loop so 1834 // insert it into the set and check whether it was already handled. 1835 auto *InnerPH = InnerL->getLoopPreheader(); 1836 assert(L.contains(InnerPH) && "Cannot contain an inner loop block " 1837 "but not contain the inner loop " 1838 "preheader!"); 1839 if (!LoopBlockSet.insert(InnerPH).second) 1840 // The only way to reach the preheader is through the loop body 1841 // itself so if it has been visited the loop is already handled. 1842 continue; 1843 1844 // Insert all of the blocks (other than those already present) into 1845 // the loop set. We expect at least the block that led us to find the 1846 // inner loop to be in the block set, but we may also have other loop 1847 // blocks if they were already enqueued as predecessors of some other 1848 // outer loop block. 1849 for (auto *InnerBB : InnerL->blocks()) { 1850 if (InnerBB == BB) { 1851 assert(LoopBlockSet.count(InnerBB) && 1852 "Block should already be in the set!"); 1853 continue; 1854 } 1855 1856 LoopBlockSet.insert(InnerBB); 1857 } 1858 1859 // Add the preheader to the worklist so we will continue past the 1860 // loop body. 1861 Worklist.push_back(InnerPH); 1862 continue; 1863 } 1864 1865 // Insert any predecessors that were in the original loop into the new 1866 // set, and if the insert is successful, add them to the worklist. 1867 for (auto *Pred : predecessors(BB)) 1868 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second) 1869 Worklist.push_back(Pred); 1870 } 1871 1872 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!"); 1873 1874 // We've found all the blocks participating in the loop, return our completed 1875 // set. 1876 return LoopBlockSet; 1877 } 1878 1879 /// Rebuild a loop after unswitching removes some subset of blocks and edges. 1880 /// 1881 /// The removal may have removed some child loops entirely but cannot have 1882 /// disturbed any remaining child loops. However, they may need to be hoisted 1883 /// to the parent loop (or to be top-level loops). The original loop may be 1884 /// completely removed. 1885 /// 1886 /// The sibling loops resulting from this update are returned. If the original 1887 /// loop remains a valid loop, it will be the first entry in this list with all 1888 /// of the newly sibling loops following it. 1889 /// 1890 /// Returns true if the loop remains a loop after unswitching, and false if it 1891 /// is no longer a loop after unswitching (and should not continue to be 1892 /// referenced). 1893 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1894 LoopInfo &LI, 1895 SmallVectorImpl<Loop *> &HoistedLoops, 1896 ScalarEvolution *SE) { 1897 auto *PH = L.getLoopPreheader(); 1898 1899 // Compute the actual parent loop from the exit blocks. Because we may have 1900 // pruned some exits the loop may be different from the original parent. 1901 Loop *ParentL = nullptr; 1902 SmallVector<Loop *, 4> ExitLoops; 1903 SmallVector<BasicBlock *, 4> ExitsInLoops; 1904 ExitsInLoops.reserve(ExitBlocks.size()); 1905 for (auto *ExitBB : ExitBlocks) 1906 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1907 ExitLoops.push_back(ExitL); 1908 ExitsInLoops.push_back(ExitBB); 1909 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1910 ParentL = ExitL; 1911 } 1912 1913 // Recompute the blocks participating in this loop. This may be empty if it 1914 // is no longer a loop. 1915 auto LoopBlockSet = recomputeLoopBlockSet(L, LI); 1916 1917 // If we still have a loop, we need to re-set the loop's parent as the exit 1918 // block set changing may have moved it within the loop nest. Note that this 1919 // can only happen when this loop has a parent as it can only hoist the loop 1920 // *up* the nest. 1921 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) { 1922 // Remove this loop's (original) blocks from all of the intervening loops. 1923 for (Loop *IL = L.getParentLoop(); IL != ParentL; 1924 IL = IL->getParentLoop()) { 1925 IL->getBlocksSet().erase(PH); 1926 for (auto *BB : L.blocks()) 1927 IL->getBlocksSet().erase(BB); 1928 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) { 1929 return BB == PH || L.contains(BB); 1930 }); 1931 } 1932 1933 LI.changeLoopFor(PH, ParentL); 1934 L.getParentLoop()->removeChildLoop(&L); 1935 if (ParentL) 1936 ParentL->addChildLoop(&L); 1937 else 1938 LI.addTopLevelLoop(&L); 1939 } 1940 1941 // Now we update all the blocks which are no longer within the loop. 1942 auto &Blocks = L.getBlocksVector(); 1943 auto BlocksSplitI = 1944 LoopBlockSet.empty() 1945 ? Blocks.begin() 1946 : std::stable_partition( 1947 Blocks.begin(), Blocks.end(), 1948 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); }); 1949 1950 // Before we erase the list of unlooped blocks, build a set of them. 1951 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end()); 1952 if (LoopBlockSet.empty()) 1953 UnloopedBlocks.insert(PH); 1954 1955 // Now erase these blocks from the loop. 1956 for (auto *BB : make_range(BlocksSplitI, Blocks.end())) 1957 L.getBlocksSet().erase(BB); 1958 Blocks.erase(BlocksSplitI, Blocks.end()); 1959 1960 // Sort the exits in ascending loop depth, we'll work backwards across these 1961 // to process them inside out. 1962 llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { 1963 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS); 1964 }); 1965 1966 // We'll build up a set for each exit loop. 1967 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks; 1968 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop. 1969 1970 auto RemoveUnloopedBlocksFromLoop = 1971 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) { 1972 for (auto *BB : UnloopedBlocks) 1973 L.getBlocksSet().erase(BB); 1974 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) { 1975 return UnloopedBlocks.count(BB); 1976 }); 1977 }; 1978 1979 SmallVector<BasicBlock *, 16> Worklist; 1980 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) { 1981 assert(Worklist.empty() && "Didn't clear worklist!"); 1982 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!"); 1983 1984 // Grab the next exit block, in decreasing loop depth order. 1985 BasicBlock *ExitBB = ExitsInLoops.pop_back_val(); 1986 Loop &ExitL = *LI.getLoopFor(ExitBB); 1987 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!"); 1988 1989 // Erase all of the unlooped blocks from the loops between the previous 1990 // exit loop and this exit loop. This works because the ExitInLoops list is 1991 // sorted in increasing order of loop depth and thus we visit loops in 1992 // decreasing order of loop depth. 1993 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop()) 1994 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1995 1996 // Walk the CFG back until we hit the cloned PH adding everything reachable 1997 // and in the unlooped set to this exit block's loop. 1998 Worklist.push_back(ExitBB); 1999 do { 2000 BasicBlock *BB = Worklist.pop_back_val(); 2001 // We can stop recursing at the cloned preheader (if we get there). 2002 if (BB == PH) 2003 continue; 2004 2005 for (BasicBlock *PredBB : predecessors(BB)) { 2006 // If this pred has already been moved to our set or is part of some 2007 // (inner) loop, no update needed. 2008 if (!UnloopedBlocks.erase(PredBB)) { 2009 assert((NewExitLoopBlocks.count(PredBB) || 2010 ExitL.contains(LI.getLoopFor(PredBB))) && 2011 "Predecessor not in a nested loop (or already visited)!"); 2012 continue; 2013 } 2014 2015 // We just insert into the loop set here. We'll add these blocks to the 2016 // exit loop after we build up the set in a deterministic order rather 2017 // than the predecessor-influenced visit order. 2018 bool Inserted = NewExitLoopBlocks.insert(PredBB).second; 2019 (void)Inserted; 2020 assert(Inserted && "Should only visit an unlooped block once!"); 2021 2022 // And recurse through to its predecessors. 2023 Worklist.push_back(PredBB); 2024 } 2025 } while (!Worklist.empty()); 2026 2027 // If blocks in this exit loop were directly part of the original loop (as 2028 // opposed to a child loop) update the map to point to this exit loop. This 2029 // just updates a map and so the fact that the order is unstable is fine. 2030 for (auto *BB : NewExitLoopBlocks) 2031 if (Loop *BBL = LI.getLoopFor(BB)) 2032 if (BBL == &L || !L.contains(BBL)) 2033 LI.changeLoopFor(BB, &ExitL); 2034 2035 // We will remove the remaining unlooped blocks from this loop in the next 2036 // iteration or below. 2037 NewExitLoopBlocks.clear(); 2038 } 2039 2040 // Any remaining unlooped blocks are no longer part of any loop unless they 2041 // are part of some child loop. 2042 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop()) 2043 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 2044 for (auto *BB : UnloopedBlocks) 2045 if (Loop *BBL = LI.getLoopFor(BB)) 2046 if (BBL == &L || !L.contains(BBL)) 2047 LI.changeLoopFor(BB, nullptr); 2048 2049 // Sink all the child loops whose headers are no longer in the loop set to 2050 // the parent (or to be top level loops). We reach into the loop and directly 2051 // update its subloop vector to make this batch update efficient. 2052 auto &SubLoops = L.getSubLoopsVector(); 2053 auto SubLoopsSplitI = 2054 LoopBlockSet.empty() 2055 ? SubLoops.begin() 2056 : std::stable_partition( 2057 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) { 2058 return LoopBlockSet.count(SubL->getHeader()); 2059 }); 2060 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) { 2061 HoistedLoops.push_back(HoistedL); 2062 HoistedL->setParentLoop(nullptr); 2063 2064 // To compute the new parent of this hoisted loop we look at where we 2065 // placed the preheader above. We can't lookup the header itself because we 2066 // retained the mapping from the header to the hoisted loop. But the 2067 // preheader and header should have the exact same new parent computed 2068 // based on the set of exit blocks from the original loop as the preheader 2069 // is a predecessor of the header and so reached in the reverse walk. And 2070 // because the loops were all in simplified form the preheader of the 2071 // hoisted loop can't be part of some *other* loop. 2072 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader())) 2073 NewParentL->addChildLoop(HoistedL); 2074 else 2075 LI.addTopLevelLoop(HoistedL); 2076 } 2077 SubLoops.erase(SubLoopsSplitI, SubLoops.end()); 2078 2079 // Actually delete the loop if nothing remained within it. 2080 if (Blocks.empty()) { 2081 assert(SubLoops.empty() && 2082 "Failed to remove all subloops from the original loop!"); 2083 if (Loop *ParentL = L.getParentLoop()) 2084 ParentL->removeChildLoop(llvm::find(*ParentL, &L)); 2085 else 2086 LI.removeLoop(llvm::find(LI, &L)); 2087 // markLoopAsDeleted for L should be triggered by the caller (it is 2088 // typically done within postUnswitch). 2089 if (SE) 2090 SE->forgetBlockAndLoopDispositions(); 2091 LI.destroy(&L); 2092 return false; 2093 } 2094 2095 return true; 2096 } 2097 2098 /// Helper to visit a dominator subtree, invoking a callable on each node. 2099 /// 2100 /// Returning false at any point will stop walking past that node of the tree. 2101 template <typename CallableT> 2102 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) { 2103 SmallVector<DomTreeNode *, 4> DomWorklist; 2104 DomWorklist.push_back(DT[BB]); 2105 #ifndef NDEBUG 2106 SmallPtrSet<DomTreeNode *, 4> Visited; 2107 Visited.insert(DT[BB]); 2108 #endif 2109 do { 2110 DomTreeNode *N = DomWorklist.pop_back_val(); 2111 2112 // Visit this node. 2113 if (!Callable(N->getBlock())) 2114 continue; 2115 2116 // Accumulate the child nodes. 2117 for (DomTreeNode *ChildN : *N) { 2118 assert(Visited.insert(ChildN).second && 2119 "Cannot visit a node twice when walking a tree!"); 2120 DomWorklist.push_back(ChildN); 2121 } 2122 } while (!DomWorklist.empty()); 2123 } 2124 2125 void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName, 2126 bool CurrentLoopValid, bool PartiallyInvariant, 2127 bool InjectedCondition, ArrayRef<Loop *> NewLoops) { 2128 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2129 if (!NewLoops.empty()) 2130 U.addSiblingLoops(NewLoops); 2131 2132 // If the current loop remains valid, we should revisit it to catch any 2133 // other unswitch opportunities. Otherwise, we need to mark it as deleted. 2134 if (CurrentLoopValid) { 2135 if (PartiallyInvariant) { 2136 // Mark the new loop as partially unswitched, to avoid unswitching on 2137 // the same condition again. 2138 auto &Context = L.getHeader()->getContext(); 2139 MDNode *DisableUnswitchMD = MDNode::get( 2140 Context, 2141 MDString::get(Context, "llvm.loop.unswitch.partial.disable")); 2142 MDNode *NewLoopID = makePostTransformationMetadata( 2143 Context, L.getLoopID(), {"llvm.loop.unswitch.partial"}, 2144 {DisableUnswitchMD}); 2145 L.setLoopID(NewLoopID); 2146 } else if (InjectedCondition) { 2147 // Do the same for injection of invariant conditions. 2148 auto &Context = L.getHeader()->getContext(); 2149 MDNode *DisableUnswitchMD = MDNode::get( 2150 Context, 2151 MDString::get(Context, "llvm.loop.unswitch.injection.disable")); 2152 MDNode *NewLoopID = makePostTransformationMetadata( 2153 Context, L.getLoopID(), {"llvm.loop.unswitch.injection"}, 2154 {DisableUnswitchMD}); 2155 L.setLoopID(NewLoopID); 2156 } else 2157 U.revisitCurrentLoop(); 2158 } else 2159 U.markLoopAsDeleted(L, LoopName); 2160 } 2161 2162 static void unswitchNontrivialInvariants( 2163 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants, 2164 IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI, 2165 AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU, 2166 LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) { 2167 auto *ParentBB = TI.getParent(); 2168 BranchInst *BI = dyn_cast<BranchInst>(&TI); 2169 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI); 2170 2171 // Save the current loop name in a variable so that we can report it even 2172 // after it has been deleted. 2173 std::string LoopName(L.getName()); 2174 2175 // We can only unswitch switches, conditional branches with an invariant 2176 // condition, or combining invariant conditions with an instruction or 2177 // partially invariant instructions. 2178 assert((SI || (BI && BI->isConditional())) && 2179 "Can only unswitch switches and conditional branch!"); 2180 bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty(); 2181 bool FullUnswitch = 2182 SI || (skipTrivialSelect(BI->getCondition()) == Invariants[0] && 2183 !PartiallyInvariant); 2184 if (FullUnswitch) 2185 assert(Invariants.size() == 1 && 2186 "Cannot have other invariants with full unswitching!"); 2187 else 2188 assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) && 2189 "Partial unswitching requires an instruction as the condition!"); 2190 2191 if (MSSAU && VerifyMemorySSA) 2192 MSSAU->getMemorySSA()->verifyMemorySSA(); 2193 2194 // Constant and BBs tracking the cloned and continuing successor. When we are 2195 // unswitching the entire condition, this can just be trivially chosen to 2196 // unswitch towards `true`. However, when we are unswitching a set of 2197 // invariants combined with `and` or `or` or partially invariant instructions, 2198 // the combining operation determines the best direction to unswitch: we want 2199 // to unswitch the direction that will collapse the branch. 2200 bool Direction = true; 2201 int ClonedSucc = 0; 2202 if (!FullUnswitch) { 2203 Value *Cond = skipTrivialSelect(BI->getCondition()); 2204 (void)Cond; 2205 assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) || 2206 PartiallyInvariant) && 2207 "Only `or`, `and`, an `select`, partially invariant instructions " 2208 "can combine invariants being unswitched."); 2209 if (!match(Cond, m_LogicalOr())) { 2210 if (match(Cond, m_LogicalAnd()) || 2211 (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) { 2212 Direction = false; 2213 ClonedSucc = 1; 2214 } 2215 } 2216 } 2217 2218 BasicBlock *RetainedSuccBB = 2219 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest(); 2220 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs; 2221 if (BI) 2222 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc)); 2223 else 2224 for (auto Case : SI->cases()) 2225 if (Case.getCaseSuccessor() != RetainedSuccBB) 2226 UnswitchedSuccBBs.insert(Case.getCaseSuccessor()); 2227 2228 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) && 2229 "Should not unswitch the same successor we are retaining!"); 2230 2231 // The branch should be in this exact loop. Any inner loop's invariant branch 2232 // should be handled by unswitching that inner loop. The caller of this 2233 // routine should filter out any candidates that remain (but were skipped for 2234 // whatever reason). 2235 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!"); 2236 2237 // Compute the parent loop now before we start hacking on things. 2238 Loop *ParentL = L.getParentLoop(); 2239 // Get blocks in RPO order for MSSA update, before changing the CFG. 2240 LoopBlocksRPO LBRPO(&L); 2241 if (MSSAU) 2242 LBRPO.perform(&LI); 2243 2244 // Compute the outer-most loop containing one of our exit blocks. This is the 2245 // furthest up our loopnest which can be mutated, which we will use below to 2246 // update things. 2247 Loop *OuterExitL = &L; 2248 SmallVector<BasicBlock *, 4> ExitBlocks; 2249 L.getUniqueExitBlocks(ExitBlocks); 2250 for (auto *ExitBB : ExitBlocks) { 2251 // ExitBB can be an exit block for several levels in the loop nest. Make 2252 // sure we find the top most. 2253 Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI); 2254 if (!NewOuterExitL) { 2255 // We exited the entire nest with this block, so we're done. 2256 OuterExitL = nullptr; 2257 break; 2258 } 2259 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL)) 2260 OuterExitL = NewOuterExitL; 2261 } 2262 2263 // At this point, we're definitely going to unswitch something so invalidate 2264 // any cached information in ScalarEvolution for the outer most loop 2265 // containing an exit block and all nested loops. 2266 if (SE) { 2267 if (OuterExitL) 2268 SE->forgetLoop(OuterExitL); 2269 else 2270 SE->forgetTopmostLoop(&L); 2271 SE->forgetBlockAndLoopDispositions(); 2272 } 2273 2274 // If the edge from this terminator to a successor dominates that successor, 2275 // store a map from each block in its dominator subtree to it. This lets us 2276 // tell when cloning for a particular successor if a block is dominated by 2277 // some *other* successor with a single data structure. We use this to 2278 // significantly reduce cloning. 2279 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc; 2280 for (auto *SuccBB : llvm::concat<BasicBlock *const>(ArrayRef(RetainedSuccBB), 2281 UnswitchedSuccBBs)) 2282 if (SuccBB->getUniquePredecessor() || 2283 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 2284 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB); 2285 })) 2286 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) { 2287 DominatingSucc[BB] = SuccBB; 2288 return true; 2289 }); 2290 2291 // Split the preheader, so that we know that there is a safe place to insert 2292 // the conditional branch. We will change the preheader to have a conditional 2293 // branch on LoopCond. The original preheader will become the split point 2294 // between the unswitched versions, and we will have a new preheader for the 2295 // original loop. 2296 BasicBlock *SplitBB = L.getLoopPreheader(); 2297 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU); 2298 2299 // Keep track of the dominator tree updates needed. 2300 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 2301 2302 // Clone the loop for each unswitched successor. 2303 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; 2304 VMaps.reserve(UnswitchedSuccBBs.size()); 2305 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs; 2306 for (auto *SuccBB : UnswitchedSuccBBs) { 2307 VMaps.emplace_back(new ValueToValueMapTy()); 2308 ClonedPHs[SuccBB] = buildClonedLoopBlocks( 2309 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB, 2310 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE); 2311 } 2312 2313 // Drop metadata if we may break its semantics by moving this instr into the 2314 // split block. 2315 if (TI.getMetadata(LLVMContext::MD_make_implicit)) { 2316 if (DropNonTrivialImplicitNullChecks) 2317 // Do not spend time trying to understand if we can keep it, just drop it 2318 // to save compile time. 2319 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr); 2320 else { 2321 // It is only legal to preserve make.implicit metadata if we are 2322 // guaranteed no reach implicit null check after following this branch. 2323 ICFLoopSafetyInfo SafetyInfo; 2324 SafetyInfo.computeLoopSafetyInfo(&L); 2325 if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L)) 2326 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr); 2327 } 2328 } 2329 2330 // The stitching of the branched code back together depends on whether we're 2331 // doing full unswitching or not with the exception that we always want to 2332 // nuke the initial terminator placed in the split block. 2333 SplitBB->getTerminator()->eraseFromParent(); 2334 if (FullUnswitch) { 2335 // Splice the terminator from the original loop and rewrite its 2336 // successors. 2337 TI.moveBefore(*SplitBB, SplitBB->end()); 2338 2339 // Keep a clone of the terminator for MSSA updates. 2340 Instruction *NewTI = TI.clone(); 2341 NewTI->insertInto(ParentBB, ParentBB->end()); 2342 2343 // First wire up the moved terminator to the preheaders. 2344 if (BI) { 2345 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2346 BI->setSuccessor(ClonedSucc, ClonedPH); 2347 BI->setSuccessor(1 - ClonedSucc, LoopPH); 2348 Value *Cond = skipTrivialSelect(BI->getCondition()); 2349 if (InsertFreeze) 2350 Cond = new FreezeInst( 2351 Cond, Cond->getName() + ".fr", BI); 2352 BI->setCondition(Cond); 2353 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2354 } else { 2355 assert(SI && "Must either be a branch or switch!"); 2356 2357 // Walk the cases and directly update their successors. 2358 assert(SI->getDefaultDest() == RetainedSuccBB && 2359 "Not retaining default successor!"); 2360 SI->setDefaultDest(LoopPH); 2361 for (const auto &Case : SI->cases()) 2362 if (Case.getCaseSuccessor() == RetainedSuccBB) 2363 Case.setSuccessor(LoopPH); 2364 else 2365 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second); 2366 2367 if (InsertFreeze) 2368 SI->setCondition(new FreezeInst( 2369 SI->getCondition(), SI->getCondition()->getName() + ".fr", SI)); 2370 2371 // We need to use the set to populate domtree updates as even when there 2372 // are multiple cases pointing at the same successor we only want to 2373 // remove and insert one edge in the domtree. 2374 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2375 DTUpdates.push_back( 2376 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second}); 2377 } 2378 2379 if (MSSAU) { 2380 DT.applyUpdates(DTUpdates); 2381 DTUpdates.clear(); 2382 2383 // Remove all but one edge to the retained block and all unswitched 2384 // blocks. This is to avoid having duplicate entries in the cloned Phis, 2385 // when we know we only keep a single edge for each case. 2386 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB); 2387 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2388 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB); 2389 2390 for (auto &VMap : VMaps) 2391 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, 2392 /*IgnoreIncomingWithNoClones=*/true); 2393 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); 2394 2395 // Remove all edges to unswitched blocks. 2396 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2397 MSSAU->removeEdge(ParentBB, SuccBB); 2398 } 2399 2400 // Now unhook the successor relationship as we'll be replacing 2401 // the terminator with a direct branch. This is much simpler for branches 2402 // than switches so we handle those first. 2403 if (BI) { 2404 // Remove the parent as a predecessor of the unswitched successor. 2405 assert(UnswitchedSuccBBs.size() == 1 && 2406 "Only one possible unswitched block for a branch!"); 2407 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin(); 2408 UnswitchedSuccBB->removePredecessor(ParentBB, 2409 /*KeepOneInputPHIs*/ true); 2410 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB}); 2411 } else { 2412 // Note that we actually want to remove the parent block as a predecessor 2413 // of *every* case successor. The case successor is either unswitched, 2414 // completely eliminating an edge from the parent to that successor, or it 2415 // is a duplicate edge to the retained successor as the retained successor 2416 // is always the default successor and as we'll replace this with a direct 2417 // branch we no longer need the duplicate entries in the PHI nodes. 2418 SwitchInst *NewSI = cast<SwitchInst>(NewTI); 2419 assert(NewSI->getDefaultDest() == RetainedSuccBB && 2420 "Not retaining default successor!"); 2421 for (const auto &Case : NewSI->cases()) 2422 Case.getCaseSuccessor()->removePredecessor( 2423 ParentBB, 2424 /*KeepOneInputPHIs*/ true); 2425 2426 // We need to use the set to populate domtree updates as even when there 2427 // are multiple cases pointing at the same successor we only want to 2428 // remove and insert one edge in the domtree. 2429 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2430 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB}); 2431 } 2432 2433 // After MSSAU update, remove the cloned terminator instruction NewTI. 2434 ParentBB->getTerminator()->eraseFromParent(); 2435 2436 // Create a new unconditional branch to the continuing block (as opposed to 2437 // the one cloned). 2438 BranchInst::Create(RetainedSuccBB, ParentBB); 2439 } else { 2440 assert(BI && "Only branches have partial unswitching."); 2441 assert(UnswitchedSuccBBs.size() == 1 && 2442 "Only one possible unswitched block for a branch!"); 2443 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2444 // When doing a partial unswitch, we have to do a bit more work to build up 2445 // the branch in the split block. 2446 if (PartiallyInvariant) 2447 buildPartialInvariantUnswitchConditionalBranch( 2448 *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU); 2449 else { 2450 buildPartialUnswitchConditionalBranch( 2451 *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, 2452 FreezeLoopUnswitchCond, BI, &AC, DT); 2453 } 2454 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2455 2456 if (MSSAU) { 2457 DT.applyUpdates(DTUpdates); 2458 DTUpdates.clear(); 2459 2460 // Perform MSSA cloning updates. 2461 for (auto &VMap : VMaps) 2462 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, 2463 /*IgnoreIncomingWithNoClones=*/true); 2464 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); 2465 } 2466 } 2467 2468 // Apply the updates accumulated above to get an up-to-date dominator tree. 2469 DT.applyUpdates(DTUpdates); 2470 2471 // Now that we have an accurate dominator tree, first delete the dead cloned 2472 // blocks so that we can accurately build any cloned loops. It is important to 2473 // not delete the blocks from the original loop yet because we still want to 2474 // reference the original loop to understand the cloned loop's structure. 2475 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU); 2476 2477 // Build the cloned loop structure itself. This may be substantially 2478 // different from the original structure due to the simplified CFG. This also 2479 // handles inserting all the cloned blocks into the correct loops. 2480 SmallVector<Loop *, 4> NonChildClonedLoops; 2481 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps) 2482 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops); 2483 2484 // Now that our cloned loops have been built, we can update the original loop. 2485 // First we delete the dead blocks from it and then we rebuild the loop 2486 // structure taking these deletions into account. 2487 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater); 2488 2489 if (MSSAU && VerifyMemorySSA) 2490 MSSAU->getMemorySSA()->verifyMemorySSA(); 2491 2492 SmallVector<Loop *, 4> HoistedLoops; 2493 bool IsStillLoop = 2494 rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE); 2495 2496 if (MSSAU && VerifyMemorySSA) 2497 MSSAU->getMemorySSA()->verifyMemorySSA(); 2498 2499 // This transformation has a high risk of corrupting the dominator tree, and 2500 // the below steps to rebuild loop structures will result in hard to debug 2501 // errors in that case so verify that the dominator tree is sane first. 2502 // FIXME: Remove this when the bugs stop showing up and rely on existing 2503 // verification steps. 2504 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2505 2506 if (BI && !PartiallyInvariant) { 2507 // If we unswitched a branch which collapses the condition to a known 2508 // constant we want to replace all the uses of the invariants within both 2509 // the original and cloned blocks. We do this here so that we can use the 2510 // now updated dominator tree to identify which side the users are on. 2511 assert(UnswitchedSuccBBs.size() == 1 && 2512 "Only one possible unswitched block for a branch!"); 2513 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2514 2515 // When considering multiple partially-unswitched invariants 2516 // we cant just go replace them with constants in both branches. 2517 // 2518 // For 'AND' we infer that true branch ("continue") means true 2519 // for each invariant operand. 2520 // For 'OR' we can infer that false branch ("continue") means false 2521 // for each invariant operand. 2522 // So it happens that for multiple-partial case we dont replace 2523 // in the unswitched branch. 2524 bool ReplaceUnswitched = 2525 FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant; 2526 2527 ConstantInt *UnswitchedReplacement = 2528 Direction ? ConstantInt::getTrue(BI->getContext()) 2529 : ConstantInt::getFalse(BI->getContext()); 2530 ConstantInt *ContinueReplacement = 2531 Direction ? ConstantInt::getFalse(BI->getContext()) 2532 : ConstantInt::getTrue(BI->getContext()); 2533 for (Value *Invariant : Invariants) { 2534 assert(!isa<Constant>(Invariant) && 2535 "Should not be replacing constant values!"); 2536 // Use make_early_inc_range here as set invalidates the iterator. 2537 for (Use &U : llvm::make_early_inc_range(Invariant->uses())) { 2538 Instruction *UserI = dyn_cast<Instruction>(U.getUser()); 2539 if (!UserI) 2540 continue; 2541 2542 // Replace it with the 'continue' side if in the main loop body, and the 2543 // unswitched if in the cloned blocks. 2544 if (DT.dominates(LoopPH, UserI->getParent())) 2545 U.set(ContinueReplacement); 2546 else if (ReplaceUnswitched && 2547 DT.dominates(ClonedPH, UserI->getParent())) 2548 U.set(UnswitchedReplacement); 2549 } 2550 } 2551 } 2552 2553 // We can change which blocks are exit blocks of all the cloned sibling 2554 // loops, the current loop, and any parent loops which shared exit blocks 2555 // with the current loop. As a consequence, we need to re-form LCSSA for 2556 // them. But we shouldn't need to re-form LCSSA for any child loops. 2557 // FIXME: This could be made more efficient by tracking which exit blocks are 2558 // new, and focusing on them, but that isn't likely to be necessary. 2559 // 2560 // In order to reasonably rebuild LCSSA we need to walk inside-out across the 2561 // loop nest and update every loop that could have had its exits changed. We 2562 // also need to cover any intervening loops. We add all of these loops to 2563 // a list and sort them by loop depth to achieve this without updating 2564 // unnecessary loops. 2565 auto UpdateLoop = [&](Loop &UpdateL) { 2566 #ifndef NDEBUG 2567 UpdateL.verifyLoop(); 2568 for (Loop *ChildL : UpdateL) { 2569 ChildL->verifyLoop(); 2570 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) && 2571 "Perturbed a child loop's LCSSA form!"); 2572 } 2573 #endif 2574 // First build LCSSA for this loop so that we can preserve it when 2575 // forming dedicated exits. We don't want to perturb some other loop's 2576 // LCSSA while doing that CFG edit. 2577 formLCSSA(UpdateL, DT, &LI, SE); 2578 2579 // For loops reached by this loop's original exit blocks we may 2580 // introduced new, non-dedicated exits. At least try to re-form dedicated 2581 // exits for these loops. This may fail if they couldn't have dedicated 2582 // exits to start with. 2583 formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true); 2584 }; 2585 2586 // For non-child cloned loops and hoisted loops, we just need to update LCSSA 2587 // and we can do it in any order as they don't nest relative to each other. 2588 // 2589 // Also check if any of the loops we have updated have become top-level loops 2590 // as that will necessitate widening the outer loop scope. 2591 for (Loop *UpdatedL : 2592 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) { 2593 UpdateLoop(*UpdatedL); 2594 if (UpdatedL->isOutermost()) 2595 OuterExitL = nullptr; 2596 } 2597 if (IsStillLoop) { 2598 UpdateLoop(L); 2599 if (L.isOutermost()) 2600 OuterExitL = nullptr; 2601 } 2602 2603 // If the original loop had exit blocks, walk up through the outer most loop 2604 // of those exit blocks to update LCSSA and form updated dedicated exits. 2605 if (OuterExitL != &L) 2606 for (Loop *OuterL = ParentL; OuterL != OuterExitL; 2607 OuterL = OuterL->getParentLoop()) 2608 UpdateLoop(*OuterL); 2609 2610 #ifndef NDEBUG 2611 // Verify the entire loop structure to catch any incorrect updates before we 2612 // progress in the pass pipeline. 2613 LI.verify(DT); 2614 #endif 2615 2616 // Now that we've unswitched something, make callbacks to report the changes. 2617 // For that we need to merge together the updated loops and the cloned loops 2618 // and check whether the original loop survived. 2619 SmallVector<Loop *, 4> SibLoops; 2620 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) 2621 if (UpdatedL->getParentLoop() == ParentL) 2622 SibLoops.push_back(UpdatedL); 2623 postUnswitch(L, LoopUpdater, LoopName, IsStillLoop, PartiallyInvariant, 2624 InjectedCondition, SibLoops); 2625 2626 if (MSSAU && VerifyMemorySSA) 2627 MSSAU->getMemorySSA()->verifyMemorySSA(); 2628 2629 if (BI) 2630 ++NumBranches; 2631 else 2632 ++NumSwitches; 2633 } 2634 2635 /// Recursively compute the cost of a dominator subtree based on the per-block 2636 /// cost map provided. 2637 /// 2638 /// The recursive computation is memozied into the provided DT-indexed cost map 2639 /// to allow querying it for most nodes in the domtree without it becoming 2640 /// quadratic. 2641 static InstructionCost computeDomSubtreeCost( 2642 DomTreeNode &N, 2643 const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap, 2644 SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) { 2645 // Don't accumulate cost (or recurse through) blocks not in our block cost 2646 // map and thus not part of the duplication cost being considered. 2647 auto BBCostIt = BBCostMap.find(N.getBlock()); 2648 if (BBCostIt == BBCostMap.end()) 2649 return 0; 2650 2651 // Lookup this node to see if we already computed its cost. 2652 auto DTCostIt = DTCostMap.find(&N); 2653 if (DTCostIt != DTCostMap.end()) 2654 return DTCostIt->second; 2655 2656 // If not, we have to compute it. We can't use insert above and update 2657 // because computing the cost may insert more things into the map. 2658 InstructionCost Cost = std::accumulate( 2659 N.begin(), N.end(), BBCostIt->second, 2660 [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost { 2661 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap); 2662 }); 2663 bool Inserted = DTCostMap.insert({&N, Cost}).second; 2664 (void)Inserted; 2665 assert(Inserted && "Should not insert a node while visiting children!"); 2666 return Cost; 2667 } 2668 2669 /// Turns a select instruction into implicit control flow branch, 2670 /// making the following replacement: 2671 /// 2672 /// head: 2673 /// --code before select-- 2674 /// select %cond, %trueval, %falseval 2675 /// --code after select-- 2676 /// 2677 /// into 2678 /// 2679 /// head: 2680 /// --code before select-- 2681 /// br i1 %cond, label %then, label %tail 2682 /// 2683 /// then: 2684 /// br %tail 2685 /// 2686 /// tail: 2687 /// phi [ %trueval, %then ], [ %falseval, %head] 2688 /// unreachable 2689 /// 2690 /// It also makes all relevant DT and LI updates, so that all structures are in 2691 /// valid state after this transform. 2692 static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT, 2693 LoopInfo &LI, MemorySSAUpdater *MSSAU, 2694 AssumptionCache *AC) { 2695 LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n"); 2696 BasicBlock *HeadBB = SI->getParent(); 2697 2698 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 2699 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false, 2700 SI->getMetadata(LLVMContext::MD_prof), &DTU, &LI); 2701 auto *CondBr = cast<BranchInst>(HeadBB->getTerminator()); 2702 BasicBlock *ThenBB = CondBr->getSuccessor(0), 2703 *TailBB = CondBr->getSuccessor(1); 2704 if (MSSAU) 2705 MSSAU->moveAllAfterSpliceBlocks(HeadBB, TailBB, SI); 2706 2707 PHINode *Phi = PHINode::Create(SI->getType(), 2, "unswitched.select", SI); 2708 Phi->addIncoming(SI->getTrueValue(), ThenBB); 2709 Phi->addIncoming(SI->getFalseValue(), HeadBB); 2710 SI->replaceAllUsesWith(Phi); 2711 SI->eraseFromParent(); 2712 2713 if (MSSAU && VerifyMemorySSA) 2714 MSSAU->getMemorySSA()->verifyMemorySSA(); 2715 2716 ++NumSelects; 2717 return CondBr; 2718 } 2719 2720 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch, 2721 /// making the following replacement: 2722 /// 2723 /// --code before guard-- 2724 /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ] 2725 /// --code after guard-- 2726 /// 2727 /// into 2728 /// 2729 /// --code before guard-- 2730 /// br i1 %cond, label %guarded, label %deopt 2731 /// 2732 /// guarded: 2733 /// --code after guard-- 2734 /// 2735 /// deopt: 2736 /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ] 2737 /// unreachable 2738 /// 2739 /// It also makes all relevant DT and LI updates, so that all structures are in 2740 /// valid state after this transform. 2741 static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L, 2742 DominatorTree &DT, LoopInfo &LI, 2743 MemorySSAUpdater *MSSAU) { 2744 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 2745 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n"); 2746 BasicBlock *CheckBB = GI->getParent(); 2747 2748 if (MSSAU && VerifyMemorySSA) 2749 MSSAU->getMemorySSA()->verifyMemorySSA(); 2750 2751 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 2752 Instruction *DeoptBlockTerm = 2753 SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true, 2754 GI->getMetadata(LLVMContext::MD_prof), &DTU, &LI); 2755 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator()); 2756 // SplitBlockAndInsertIfThen inserts control flow that branches to 2757 // DeoptBlockTerm if the condition is true. We want the opposite. 2758 CheckBI->swapSuccessors(); 2759 2760 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0); 2761 GuardedBlock->setName("guarded"); 2762 CheckBI->getSuccessor(1)->setName("deopt"); 2763 BasicBlock *DeoptBlock = CheckBI->getSuccessor(1); 2764 2765 if (MSSAU) 2766 MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI); 2767 2768 GI->moveBefore(DeoptBlockTerm); 2769 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext())); 2770 2771 if (MSSAU) { 2772 MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI)); 2773 MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator); 2774 if (VerifyMemorySSA) 2775 MSSAU->getMemorySSA()->verifyMemorySSA(); 2776 } 2777 2778 if (VerifyLoopInfo) 2779 LI.verify(DT); 2780 ++NumGuards; 2781 return CheckBI; 2782 } 2783 2784 /// Cost multiplier is a way to limit potentially exponential behavior 2785 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch 2786 /// candidates available. Also accounting for the number of "sibling" loops with 2787 /// the idea to account for previous unswitches that already happened on this 2788 /// cluster of loops. There was an attempt to keep this formula simple, 2789 /// just enough to limit the worst case behavior. Even if it is not that simple 2790 /// now it is still not an attempt to provide a detailed heuristic size 2791 /// prediction. 2792 /// 2793 /// TODO: Make a proper accounting of "explosion" effect for all kinds of 2794 /// unswitch candidates, making adequate predictions instead of wild guesses. 2795 /// That requires knowing not just the number of "remaining" candidates but 2796 /// also costs of unswitching for each of these candidates. 2797 static int CalculateUnswitchCostMultiplier( 2798 const Instruction &TI, const Loop &L, const LoopInfo &LI, 2799 const DominatorTree &DT, 2800 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) { 2801 2802 // Guards and other exiting conditions do not contribute to exponential 2803 // explosion as soon as they dominate the latch (otherwise there might be 2804 // another path to the latch remaining that does not allow to eliminate the 2805 // loop copy on unswitch). 2806 const BasicBlock *Latch = L.getLoopLatch(); 2807 const BasicBlock *CondBlock = TI.getParent(); 2808 if (DT.dominates(CondBlock, Latch) && 2809 (isGuard(&TI) || 2810 (TI.isTerminator() && 2811 llvm::count_if(successors(&TI), [&L](const BasicBlock *SuccBB) { 2812 return L.contains(SuccBB); 2813 }) <= 1))) { 2814 NumCostMultiplierSkipped++; 2815 return 1; 2816 } 2817 2818 auto *ParentL = L.getParentLoop(); 2819 int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size() 2820 : std::distance(LI.begin(), LI.end())); 2821 // Count amount of clones that all the candidates might cause during 2822 // unswitching. Branch/guard/select counts as 1, switch counts as log2 of its 2823 // cases. 2824 int UnswitchedClones = 0; 2825 for (const auto &Candidate : UnswitchCandidates) { 2826 const Instruction *CI = Candidate.TI; 2827 const BasicBlock *CondBlock = CI->getParent(); 2828 bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch); 2829 if (isa<SelectInst>(CI)) { 2830 UnswitchedClones++; 2831 continue; 2832 } 2833 if (isGuard(CI)) { 2834 if (!SkipExitingSuccessors) 2835 UnswitchedClones++; 2836 continue; 2837 } 2838 int NonExitingSuccessors = 2839 llvm::count_if(successors(CondBlock), 2840 [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) { 2841 return !SkipExitingSuccessors || L.contains(SuccBB); 2842 }); 2843 UnswitchedClones += Log2_32(NonExitingSuccessors); 2844 } 2845 2846 // Ignore up to the "unscaled candidates" number of unswitch candidates 2847 // when calculating the power-of-two scaling of the cost. The main idea 2848 // with this control is to allow a small number of unswitches to happen 2849 // and rely more on siblings multiplier (see below) when the number 2850 // of candidates is small. 2851 unsigned ClonesPower = 2852 std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0); 2853 2854 // Allowing top-level loops to spread a bit more than nested ones. 2855 int SiblingsMultiplier = 2856 std::max((ParentL ? SiblingsCount 2857 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv), 2858 1); 2859 // Compute the cost multiplier in a way that won't overflow by saturating 2860 // at an upper bound. 2861 int CostMultiplier; 2862 if (ClonesPower > Log2_32(UnswitchThreshold) || 2863 SiblingsMultiplier > UnswitchThreshold) 2864 CostMultiplier = UnswitchThreshold; 2865 else 2866 CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower), 2867 (int)UnswitchThreshold); 2868 2869 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier 2870 << " (siblings " << SiblingsMultiplier << " * clones " 2871 << (1 << ClonesPower) << ")" 2872 << " for unswitch candidate: " << TI << "\n"); 2873 return CostMultiplier; 2874 } 2875 2876 static bool collectUnswitchCandidates( 2877 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, 2878 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, 2879 const Loop &L, const LoopInfo &LI, AAResults &AA, 2880 const MemorySSAUpdater *MSSAU) { 2881 assert(UnswitchCandidates.empty() && "Should be!"); 2882 2883 auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) { 2884 Cond = skipTrivialSelect(Cond); 2885 if (isa<Constant>(Cond)) 2886 return; 2887 if (L.isLoopInvariant(Cond)) { 2888 UnswitchCandidates.push_back({I, {Cond}}); 2889 return; 2890 } 2891 if (match(Cond, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) { 2892 TinyPtrVector<Value *> Invariants = 2893 collectHomogenousInstGraphLoopInvariants( 2894 L, *static_cast<Instruction *>(Cond), LI); 2895 if (!Invariants.empty()) 2896 UnswitchCandidates.push_back({I, std::move(Invariants)}); 2897 } 2898 }; 2899 2900 // Whether or not we should also collect guards in the loop. 2901 bool CollectGuards = false; 2902 if (UnswitchGuards) { 2903 auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction( 2904 Intrinsic::getName(Intrinsic::experimental_guard)); 2905 if (GuardDecl && !GuardDecl->use_empty()) 2906 CollectGuards = true; 2907 } 2908 2909 for (auto *BB : L.blocks()) { 2910 if (LI.getLoopFor(BB) != &L) 2911 continue; 2912 2913 for (auto &I : *BB) { 2914 if (auto *SI = dyn_cast<SelectInst>(&I)) { 2915 auto *Cond = SI->getCondition(); 2916 // Do not unswitch vector selects and logical and/or selects 2917 if (Cond->getType()->isIntegerTy(1) && !SI->getType()->isIntegerTy(1)) 2918 AddUnswitchCandidatesForInst(SI, Cond); 2919 } else if (CollectGuards && isGuard(&I)) { 2920 auto *Cond = 2921 skipTrivialSelect(cast<IntrinsicInst>(&I)->getArgOperand(0)); 2922 // TODO: Support AND, OR conditions and partial unswitching. 2923 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond)) 2924 UnswitchCandidates.push_back({&I, {Cond}}); 2925 } 2926 } 2927 2928 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2929 // We can only consider fully loop-invariant switch conditions as we need 2930 // to completely eliminate the switch after unswitching. 2931 if (!isa<Constant>(SI->getCondition()) && 2932 L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor()) 2933 UnswitchCandidates.push_back({SI, {SI->getCondition()}}); 2934 continue; 2935 } 2936 2937 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 2938 if (!BI || !BI->isConditional() || 2939 BI->getSuccessor(0) == BI->getSuccessor(1)) 2940 continue; 2941 2942 AddUnswitchCandidatesForInst(BI, BI->getCondition()); 2943 } 2944 2945 if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") && 2946 !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) { 2947 return TerminatorAndInvariants.TI == L.getHeader()->getTerminator(); 2948 })) { 2949 MemorySSA *MSSA = MSSAU->getMemorySSA(); 2950 if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) { 2951 LLVM_DEBUG( 2952 dbgs() << "simple-loop-unswitch: Found partially invariant condition " 2953 << *Info->InstToDuplicate[0] << "\n"); 2954 PartialIVInfo = *Info; 2955 PartialIVCondBranch = L.getHeader()->getTerminator(); 2956 TinyPtrVector<Value *> ValsToDuplicate; 2957 llvm::append_range(ValsToDuplicate, Info->InstToDuplicate); 2958 UnswitchCandidates.push_back( 2959 {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)}); 2960 } 2961 } 2962 return !UnswitchCandidates.empty(); 2963 } 2964 2965 /// Tries to canonicalize condition described by: 2966 /// 2967 /// br (LHS pred RHS), label IfTrue, label IfFalse 2968 /// 2969 /// into its equivalent where `Pred` is something that we support for injected 2970 /// invariants (so far it is limited to ult), LHS in canonicalized form is 2971 /// non-invariant and RHS is an invariant. 2972 static void canonicalizeForInvariantConditionInjection( 2973 ICmpInst::Predicate &Pred, Value *&LHS, Value *&RHS, BasicBlock *&IfTrue, 2974 BasicBlock *&IfFalse, const Loop &L) { 2975 if (!L.contains(IfTrue)) { 2976 Pred = ICmpInst::getInversePredicate(Pred); 2977 std::swap(IfTrue, IfFalse); 2978 } 2979 2980 // Move loop-invariant argument to RHS position. 2981 if (L.isLoopInvariant(LHS)) { 2982 Pred = ICmpInst::getSwappedPredicate(Pred); 2983 std::swap(LHS, RHS); 2984 } 2985 2986 if (Pred == ICmpInst::ICMP_SGE && match(RHS, m_Zero())) { 2987 // Turn "x >=s 0" into "x <u UMIN_INT" 2988 Pred = ICmpInst::ICMP_ULT; 2989 RHS = ConstantInt::get( 2990 RHS->getContext(), 2991 APInt::getSignedMinValue(RHS->getType()->getIntegerBitWidth())); 2992 } 2993 } 2994 2995 /// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS ) 2996 /// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by 2997 /// injecting a loop-invariant condition. 2998 static bool shouldTryInjectInvariantCondition( 2999 const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS, 3000 const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) { 3001 if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS)) 3002 return false; 3003 // TODO: Support other predicates. 3004 if (Pred != ICmpInst::ICMP_ULT) 3005 return false; 3006 // TODO: Support non-loop-exiting branches? 3007 if (!L.contains(IfTrue) || L.contains(IfFalse)) 3008 return false; 3009 // FIXME: For some reason this causes problems with MSSA updates, need to 3010 // investigate why. So far, just don't unswitch latch. 3011 if (L.getHeader() == IfTrue) 3012 return false; 3013 return true; 3014 } 3015 3016 /// Returns true, if metadata on \p BI allows us to optimize branching into \p 3017 /// TakenSucc via injection of invariant conditions. The branch should be not 3018 /// enough and not previously unswitched, the information about this comes from 3019 /// the metadata. 3020 bool shouldTryInjectBasingOnMetadata(const BranchInst *BI, 3021 const BasicBlock *TakenSucc) { 3022 SmallVector<uint32_t> Weights; 3023 if (!extractBranchWeights(*BI, Weights)) 3024 return false; 3025 unsigned T = InjectInvariantConditionHotnesThreshold; 3026 BranchProbability LikelyTaken(T - 1, T); 3027 3028 assert(Weights.size() == 2 && "Unexpected profile data!"); 3029 size_t Idx = BI->getSuccessor(0) == TakenSucc ? 0 : 1; 3030 auto Num = Weights[Idx]; 3031 auto Denom = Weights[0] + Weights[1]; 3032 // Degenerate or overflowed metadata. 3033 if (Denom == 0 || Num > Denom) 3034 return false; 3035 BranchProbability ActualTaken(Num, Denom); 3036 if (LikelyTaken > ActualTaken) 3037 return false; 3038 return true; 3039 } 3040 3041 /// Materialize pending invariant condition of the given candidate into IR. The 3042 /// injected loop-invariant condition implies the original loop-variant branch 3043 /// condition, so the materialization turns 3044 /// 3045 /// loop_block: 3046 /// ... 3047 /// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc 3048 /// 3049 /// into 3050 /// 3051 /// preheader: 3052 /// %invariant_cond = LHS pred RHS 3053 /// ... 3054 /// loop_block: 3055 /// br i1 %invariant_cond, label InLoopSucc, label OriginalCheck 3056 /// OriginalCheck: 3057 /// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc 3058 /// ... 3059 static NonTrivialUnswitchCandidate 3060 injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L, 3061 DominatorTree &DT, LoopInfo &LI, 3062 AssumptionCache &AC, MemorySSAUpdater *MSSAU) { 3063 assert(Candidate.hasPendingInjection() && "Nothing to inject!"); 3064 BasicBlock *Preheader = L.getLoopPreheader(); 3065 assert(Preheader && "Loop is not in simplified form?"); 3066 assert(LI.getLoopFor(Candidate.TI->getParent()) == &L && 3067 "Unswitching branch of inner loop!"); 3068 3069 auto Pred = Candidate.PendingInjection->Pred; 3070 auto *LHS = Candidate.PendingInjection->LHS; 3071 auto *RHS = Candidate.PendingInjection->RHS; 3072 auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc; 3073 auto *TI = cast<BranchInst>(Candidate.TI); 3074 auto *BB = Candidate.TI->getParent(); 3075 auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(0) ? TI->getSuccessor(1) 3076 : TI->getSuccessor(0); 3077 // FIXME: Remove this once limitation on successors is lifted. 3078 assert(L.contains(InLoopSucc) && "Not supported yet!"); 3079 assert(!L.contains(OutOfLoopSucc) && "Not supported yet!"); 3080 auto &Ctx = BB->getContext(); 3081 3082 IRBuilder<> Builder(Preheader->getTerminator()); 3083 assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!"); 3084 if (LHS->getType() != RHS->getType()) { 3085 if (LHS->getType()->getIntegerBitWidth() < 3086 RHS->getType()->getIntegerBitWidth()) 3087 LHS = Builder.CreateZExt(LHS, RHS->getType(), LHS->getName() + ".wide"); 3088 else 3089 RHS = Builder.CreateZExt(RHS, LHS->getType(), RHS->getName() + ".wide"); 3090 } 3091 // Do not use builder here: CreateICmp may simplify this into a constant and 3092 // unswitching will break. Better optimize it away later. 3093 auto *InjectedCond = 3094 ICmpInst::Create(Instruction::ICmp, Pred, LHS, RHS, "injected.cond", 3095 Preheader->getTerminator()); 3096 3097 BasicBlock *CheckBlock = BasicBlock::Create(Ctx, BB->getName() + ".check", 3098 BB->getParent(), InLoopSucc); 3099 Builder.SetInsertPoint(TI); 3100 auto *InvariantBr = 3101 Builder.CreateCondBr(InjectedCond, InLoopSucc, CheckBlock); 3102 3103 Builder.SetInsertPoint(CheckBlock); 3104 Builder.CreateCondBr(TI->getCondition(), TI->getSuccessor(0), 3105 TI->getSuccessor(1)); 3106 TI->eraseFromParent(); 3107 3108 // Fixup phis. 3109 for (auto &I : *InLoopSucc) { 3110 auto *PN = dyn_cast<PHINode>(&I); 3111 if (!PN) 3112 break; 3113 auto *Inc = PN->getIncomingValueForBlock(BB); 3114 PN->addIncoming(Inc, CheckBlock); 3115 } 3116 OutOfLoopSucc->replacePhiUsesWith(BB, CheckBlock); 3117 3118 SmallVector<DominatorTree::UpdateType, 4> DTUpdates = { 3119 { DominatorTree::Insert, BB, CheckBlock }, 3120 { DominatorTree::Insert, CheckBlock, InLoopSucc }, 3121 { DominatorTree::Insert, CheckBlock, OutOfLoopSucc }, 3122 { DominatorTree::Delete, BB, OutOfLoopSucc } 3123 }; 3124 3125 DT.applyUpdates(DTUpdates); 3126 if (MSSAU) 3127 MSSAU->applyUpdates(DTUpdates, DT); 3128 L.addBasicBlockToLoop(CheckBlock, LI); 3129 3130 #ifndef NDEBUG 3131 DT.verify(); 3132 LI.verify(DT); 3133 if (MSSAU && VerifyMemorySSA) 3134 MSSAU->getMemorySSA()->verifyMemorySSA(); 3135 #endif 3136 3137 // TODO: In fact, cost of unswitching a new invariant candidate is *slightly* 3138 // higher because we have just inserted a new block. Need to think how to 3139 // adjust the cost of injected candidates when it was first computed. 3140 LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr 3141 << " and considering it for unswitching."); 3142 ++NumInvariantConditionsInjected; 3143 return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond }, 3144 Candidate.Cost); 3145 } 3146 3147 /// Given chain of loop branch conditions looking like: 3148 /// br (Variant < Invariant1) 3149 /// br (Variant < Invariant2) 3150 /// br (Variant < Invariant3) 3151 /// ... 3152 /// collect set of invariant conditions on which we want to unswitch, which 3153 /// look like: 3154 /// Invariant1 <= Invariant2 3155 /// Invariant2 <= Invariant3 3156 /// ... 3157 /// Though they might not immediately exist in the IR, we can still inject them. 3158 static bool insertCandidatesWithPendingInjections( 3159 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L, 3160 ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares, 3161 const DominatorTree &DT) { 3162 3163 assert(ICmpInst::isRelational(Pred)); 3164 assert(ICmpInst::isStrictPredicate(Pred)); 3165 if (Compares.size() < 2) 3166 return false; 3167 ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(Pred); 3168 for (auto Prev = Compares.begin(), Next = Compares.begin() + 1; 3169 Next != Compares.end(); ++Prev, ++Next) { 3170 Value *LHS = Next->Invariant; 3171 Value *RHS = Prev->Invariant; 3172 BasicBlock *InLoopSucc = Prev->InLoopSucc; 3173 InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc); 3174 NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS }, 3175 std::nullopt, std::move(ToInject)); 3176 UnswitchCandidates.push_back(std::move(Candidate)); 3177 } 3178 return true; 3179 } 3180 3181 /// Collect unswitch candidates by invariant conditions that are not immediately 3182 /// present in the loop. However, they can be injected into the code if we 3183 /// decide it's profitable. 3184 /// An example of such conditions is following: 3185 /// 3186 /// for (...) { 3187 /// x = load ... 3188 /// if (! x <u C1) break; 3189 /// if (! x <u C2) break; 3190 /// <do something> 3191 /// } 3192 /// 3193 /// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <= 3194 /// C2" automatically implies "x <u C2", so we can get rid of one of 3195 /// loop-variant checks in unswitched loop version. 3196 static bool collectUnswitchCandidatesWithInjections( 3197 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, 3198 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L, 3199 const DominatorTree &DT, const LoopInfo &LI, AAResults &AA, 3200 const MemorySSAUpdater *MSSAU) { 3201 if (!InjectInvariantConditions) 3202 return false; 3203 3204 if (!DT.isReachableFromEntry(L.getHeader())) 3205 return false; 3206 auto *Latch = L.getLoopLatch(); 3207 // Need to have a single latch and a preheader. 3208 if (!Latch) 3209 return false; 3210 assert(L.getLoopPreheader() && "Must have a preheader!"); 3211 3212 DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT; 3213 // Traverse the conditions that dominate latch (and therefore dominate each 3214 // other). 3215 for (auto *DTN = DT.getNode(Latch); L.contains(DTN->getBlock()); 3216 DTN = DTN->getIDom()) { 3217 ICmpInst::Predicate Pred; 3218 Value *LHS = nullptr, *RHS = nullptr; 3219 BasicBlock *IfTrue = nullptr, *IfFalse = nullptr; 3220 auto *BB = DTN->getBlock(); 3221 // Ignore inner loops. 3222 if (LI.getLoopFor(BB) != &L) 3223 continue; 3224 auto *Term = BB->getTerminator(); 3225 if (!match(Term, m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)), 3226 m_BasicBlock(IfTrue), m_BasicBlock(IfFalse)))) 3227 continue; 3228 if (!LHS->getType()->isIntegerTy()) 3229 continue; 3230 canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse, 3231 L); 3232 if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L)) 3233 continue; 3234 if (!shouldTryInjectBasingOnMetadata(cast<BranchInst>(Term), IfTrue)) 3235 continue; 3236 // Strip ZEXT for unsigned predicate. 3237 // TODO: once signed predicates are supported, also strip SEXT. 3238 CompareDesc Desc(cast<BranchInst>(Term), RHS, IfTrue); 3239 while (auto *Zext = dyn_cast<ZExtInst>(LHS)) 3240 LHS = Zext->getOperand(0); 3241 CandidatesULT[LHS].push_back(Desc); 3242 } 3243 3244 bool Found = false; 3245 for (auto &It : CandidatesULT) 3246 Found |= insertCandidatesWithPendingInjections( 3247 UnswitchCandidates, L, ICmpInst::ICMP_ULT, It.second, DT); 3248 return Found; 3249 } 3250 3251 static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) { 3252 if (!L.isSafeToClone()) 3253 return false; 3254 for (auto *BB : L.blocks()) 3255 for (auto &I : *BB) { 3256 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB)) 3257 return false; 3258 if (auto *CB = dyn_cast<CallBase>(&I)) { 3259 assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone()."); 3260 if (CB->isConvergent()) 3261 return false; 3262 } 3263 } 3264 3265 // Check if there are irreducible CFG cycles in this loop. If so, we cannot 3266 // easily unswitch non-trivial edges out of the loop. Doing so might turn the 3267 // irreducible control flow into reducible control flow and introduce new 3268 // loops "out of thin air". If we ever discover important use cases for doing 3269 // this, we can add support to loop unswitch, but it is a lot of complexity 3270 // for what seems little or no real world benefit. 3271 LoopBlocksRPO RPOT(&L); 3272 RPOT.perform(&LI); 3273 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 3274 return false; 3275 3276 SmallVector<BasicBlock *, 4> ExitBlocks; 3277 L.getUniqueExitBlocks(ExitBlocks); 3278 // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch 3279 // instruction as we don't know how to split those exit blocks. 3280 // FIXME: We should teach SplitBlock to handle this and remove this 3281 // restriction. 3282 for (auto *ExitBB : ExitBlocks) { 3283 auto *I = ExitBB->getFirstNonPHI(); 3284 if (isa<CleanupPadInst>(I) || isa<CatchSwitchInst>(I)) { 3285 LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch " 3286 "in exit block\n"); 3287 return false; 3288 } 3289 } 3290 3291 return true; 3292 } 3293 3294 static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate( 3295 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L, 3296 const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC, 3297 const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) { 3298 // Given that unswitching these terminators will require duplicating parts of 3299 // the loop, so we need to be able to model that cost. Compute the ephemeral 3300 // values and set up a data structure to hold per-BB costs. We cache each 3301 // block's cost so that we don't recompute this when considering different 3302 // subsets of the loop for duplication during unswitching. 3303 SmallPtrSet<const Value *, 4> EphValues; 3304 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues); 3305 SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap; 3306 3307 // Compute the cost of each block, as well as the total loop cost. Also, bail 3308 // out if we see instructions which are incompatible with loop unswitching 3309 // (convergent, noduplicate, or cross-basic-block tokens). 3310 // FIXME: We might be able to safely handle some of these in non-duplicated 3311 // regions. 3312 TargetTransformInfo::TargetCostKind CostKind = 3313 L.getHeader()->getParent()->hasMinSize() 3314 ? TargetTransformInfo::TCK_CodeSize 3315 : TargetTransformInfo::TCK_SizeAndLatency; 3316 InstructionCost LoopCost = 0; 3317 for (auto *BB : L.blocks()) { 3318 InstructionCost Cost = 0; 3319 for (auto &I : *BB) { 3320 if (EphValues.count(&I)) 3321 continue; 3322 Cost += TTI.getInstructionCost(&I, CostKind); 3323 } 3324 assert(Cost >= 0 && "Must not have negative costs!"); 3325 LoopCost += Cost; 3326 assert(LoopCost >= 0 && "Must not have negative loop costs!"); 3327 BBCostMap[BB] = Cost; 3328 } 3329 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n"); 3330 3331 // Now we find the best candidate by searching for the one with the following 3332 // properties in order: 3333 // 3334 // 1) An unswitching cost below the threshold 3335 // 2) The smallest number of duplicated unswitch candidates (to avoid 3336 // creating redundant subsequent unswitching) 3337 // 3) The smallest cost after unswitching. 3338 // 3339 // We prioritize reducing fanout of unswitch candidates provided the cost 3340 // remains below the threshold because this has a multiplicative effect. 3341 // 3342 // This requires memoizing each dominator subtree to avoid redundant work. 3343 // 3344 // FIXME: Need to actually do the number of candidates part above. 3345 SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap; 3346 // Given a terminator which might be unswitched, computes the non-duplicated 3347 // cost for that terminator. 3348 auto ComputeUnswitchedCost = [&](Instruction &TI, 3349 bool FullUnswitch) -> InstructionCost { 3350 // Unswitching selects unswitches the entire loop. 3351 if (isa<SelectInst>(TI)) 3352 return LoopCost; 3353 3354 BasicBlock &BB = *TI.getParent(); 3355 SmallPtrSet<BasicBlock *, 4> Visited; 3356 3357 InstructionCost Cost = 0; 3358 for (BasicBlock *SuccBB : successors(&BB)) { 3359 // Don't count successors more than once. 3360 if (!Visited.insert(SuccBB).second) 3361 continue; 3362 3363 // If this is a partial unswitch candidate, then it must be a conditional 3364 // branch with a condition of either `or`, `and`, their corresponding 3365 // select forms or partially invariant instructions. In that case, one of 3366 // the successors is necessarily duplicated, so don't even try to remove 3367 // its cost. 3368 if (!FullUnswitch) { 3369 auto &BI = cast<BranchInst>(TI); 3370 Value *Cond = skipTrivialSelect(BI.getCondition()); 3371 if (match(Cond, m_LogicalAnd())) { 3372 if (SuccBB == BI.getSuccessor(1)) 3373 continue; 3374 } else if (match(Cond, m_LogicalOr())) { 3375 if (SuccBB == BI.getSuccessor(0)) 3376 continue; 3377 } else if ((PartialIVInfo.KnownValue->isOneValue() && 3378 SuccBB == BI.getSuccessor(0)) || 3379 (!PartialIVInfo.KnownValue->isOneValue() && 3380 SuccBB == BI.getSuccessor(1))) 3381 continue; 3382 } 3383 3384 // This successor's domtree will not need to be duplicated after 3385 // unswitching if the edge to the successor dominates it (and thus the 3386 // entire tree). This essentially means there is no other path into this 3387 // subtree and so it will end up live in only one clone of the loop. 3388 if (SuccBB->getUniquePredecessor() || 3389 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 3390 return PredBB == &BB || DT.dominates(SuccBB, PredBB); 3391 })) { 3392 Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap); 3393 assert(Cost <= LoopCost && 3394 "Non-duplicated cost should never exceed total loop cost!"); 3395 } 3396 } 3397 3398 // Now scale the cost by the number of unique successors minus one. We 3399 // subtract one because there is already at least one copy of the entire 3400 // loop. This is computing the new cost of unswitching a condition. 3401 // Note that guards always have 2 unique successors that are implicit and 3402 // will be materialized if we decide to unswitch it. 3403 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size(); 3404 assert(SuccessorsCount > 1 && 3405 "Cannot unswitch a condition without multiple distinct successors!"); 3406 return (LoopCost - Cost) * (SuccessorsCount - 1); 3407 }; 3408 3409 std::optional<NonTrivialUnswitchCandidate> Best; 3410 for (auto &Candidate : UnswitchCandidates) { 3411 Instruction &TI = *Candidate.TI; 3412 ArrayRef<Value *> Invariants = Candidate.Invariants; 3413 BranchInst *BI = dyn_cast<BranchInst>(&TI); 3414 bool FullUnswitch = 3415 !BI || Candidate.hasPendingInjection() || 3416 (Invariants.size() == 1 && 3417 Invariants[0] == skipTrivialSelect(BI->getCondition())); 3418 InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch); 3419 // Calculate cost multiplier which is a tool to limit potentially 3420 // exponential behavior of loop-unswitch. 3421 if (EnableUnswitchCostMultiplier) { 3422 int CostMultiplier = 3423 CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates); 3424 assert( 3425 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) && 3426 "cost multiplier needs to be in the range of 1..UnswitchThreshold"); 3427 CandidateCost *= CostMultiplier; 3428 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 3429 << " (multiplier: " << CostMultiplier << ")" 3430 << " for unswitch candidate: " << TI << "\n"); 3431 } else { 3432 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 3433 << " for unswitch candidate: " << TI << "\n"); 3434 } 3435 3436 if (!Best || CandidateCost < Best->Cost) { 3437 Best = Candidate; 3438 Best->Cost = CandidateCost; 3439 } 3440 } 3441 assert(Best && "Must be!"); 3442 return *Best; 3443 } 3444 3445 // Insert a freeze on an unswitched branch if all is true: 3446 // 1. freeze-loop-unswitch-cond option is true 3447 // 2. The branch may not execute in the loop pre-transformation. If a branch may 3448 // not execute and could cause UB, it would always cause UB if it is hoisted outside 3449 // of the loop. Insert a freeze to prevent this case. 3450 // 3. The branch condition may be poison or undef 3451 static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT, 3452 AssumptionCache &AC) { 3453 assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI)); 3454 if (!FreezeLoopUnswitchCond) 3455 return false; 3456 3457 ICFLoopSafetyInfo SafetyInfo; 3458 SafetyInfo.computeLoopSafetyInfo(&L); 3459 if (SafetyInfo.isGuaranteedToExecute(TI, &DT, &L)) 3460 return false; 3461 3462 Value *Cond; 3463 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) 3464 Cond = skipTrivialSelect(BI->getCondition()); 3465 else 3466 Cond = skipTrivialSelect(cast<SwitchInst>(&TI)->getCondition()); 3467 return !isGuaranteedNotToBeUndefOrPoison( 3468 Cond, &AC, L.getLoopPreheader()->getTerminator(), &DT); 3469 } 3470 3471 static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI, 3472 AssumptionCache &AC, AAResults &AA, 3473 TargetTransformInfo &TTI, ScalarEvolution *SE, 3474 MemorySSAUpdater *MSSAU, 3475 LPMUpdater &LoopUpdater) { 3476 // Collect all invariant conditions within this loop (as opposed to an inner 3477 // loop which would be handled when visiting that inner loop). 3478 SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates; 3479 IVConditionInfo PartialIVInfo; 3480 Instruction *PartialIVCondBranch = nullptr; 3481 collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo, 3482 PartialIVCondBranch, L, LI, AA, MSSAU); 3483 if (!findOptionMDForLoop(&L, "llvm.loop.unswitch.injection.disable")) 3484 collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo, 3485 PartialIVCondBranch, L, DT, LI, AA, 3486 MSSAU); 3487 // If we didn't find any candidates, we're done. 3488 if (UnswitchCandidates.empty()) 3489 return false; 3490 3491 LLVM_DEBUG( 3492 dbgs() << "Considering " << UnswitchCandidates.size() 3493 << " non-trivial loop invariant conditions for unswitching.\n"); 3494 3495 NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate( 3496 UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo); 3497 3498 assert(Best.TI && "Failed to find loop unswitch candidate"); 3499 assert(Best.Cost && "Failed to compute cost"); 3500 3501 if (*Best.Cost >= UnswitchThreshold) { 3502 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost 3503 << "\n"); 3504 return false; 3505 } 3506 3507 bool InjectedCondition = false; 3508 if (Best.hasPendingInjection()) { 3509 Best = injectPendingInvariantConditions(Best, L, DT, LI, AC, MSSAU); 3510 InjectedCondition = true; 3511 } 3512 assert(!Best.hasPendingInjection() && 3513 "All injections should have been done by now!"); 3514 3515 if (Best.TI != PartialIVCondBranch) 3516 PartialIVInfo.InstToDuplicate.clear(); 3517 3518 bool InsertFreeze; 3519 if (auto *SI = dyn_cast<SelectInst>(Best.TI)) { 3520 // If the best candidate is a select, turn it into a branch. Select 3521 // instructions with a poison conditional do not propagate poison, but 3522 // branching on poison causes UB. Insert a freeze on the select 3523 // conditional to prevent UB after turning the select into a branch. 3524 InsertFreeze = !isGuaranteedNotToBeUndefOrPoison( 3525 SI->getCondition(), &AC, L.getLoopPreheader()->getTerminator(), &DT); 3526 Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, &AC); 3527 } else { 3528 // If the best candidate is a guard, turn it into a branch. 3529 if (isGuard(Best.TI)) 3530 Best.TI = 3531 turnGuardIntoBranch(cast<IntrinsicInst>(Best.TI), L, DT, LI, MSSAU); 3532 InsertFreeze = shouldInsertFreeze(L, *Best.TI, DT, AC); 3533 } 3534 3535 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = " << Best.Cost 3536 << ") terminator: " << *Best.TI << "\n"); 3537 unswitchNontrivialInvariants(L, *Best.TI, Best.Invariants, PartialIVInfo, DT, 3538 LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze, 3539 InjectedCondition); 3540 return true; 3541 } 3542 3543 /// Unswitch control flow predicated on loop invariant conditions. 3544 /// 3545 /// This first hoists all branches or switches which are trivial (IE, do not 3546 /// require duplicating any part of the loop) out of the loop body. It then 3547 /// looks at other loop invariant control flows and tries to unswitch those as 3548 /// well by cloning the loop if the result is small enough. 3549 /// 3550 /// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are 3551 /// also updated based on the unswitch. The `MSSA` analysis is also updated if 3552 /// valid (i.e. its use is enabled). 3553 /// 3554 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is 3555 /// true, we will attempt to do non-trivial unswitching as well as trivial 3556 /// unswitching. 3557 /// 3558 /// The `postUnswitch` function will be run after unswitching is complete 3559 /// with information on whether or not the provided loop remains a loop and 3560 /// a list of new sibling loops created. 3561 /// 3562 /// If `SE` is non-null, we will update that analysis based on the unswitching 3563 /// done. 3564 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, 3565 AssumptionCache &AC, AAResults &AA, 3566 TargetTransformInfo &TTI, bool Trivial, 3567 bool NonTrivial, ScalarEvolution *SE, 3568 MemorySSAUpdater *MSSAU, ProfileSummaryInfo *PSI, 3569 BlockFrequencyInfo *BFI, LPMUpdater &LoopUpdater) { 3570 assert(L.isRecursivelyLCSSAForm(DT, LI) && 3571 "Loops must be in LCSSA form before unswitching."); 3572 3573 // Must be in loop simplified form: we need a preheader and dedicated exits. 3574 if (!L.isLoopSimplifyForm()) 3575 return false; 3576 3577 // Try trivial unswitch first before loop over other basic blocks in the loop. 3578 if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) { 3579 // If we unswitched successfully we will want to clean up the loop before 3580 // processing it further so just mark it as unswitched and return. 3581 postUnswitch(L, LoopUpdater, L.getName(), 3582 /*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false, 3583 /*InjectedCondition*/ false, {}); 3584 return true; 3585 } 3586 3587 const Function *F = L.getHeader()->getParent(); 3588 3589 // Check whether we should continue with non-trivial conditions. 3590 // EnableNonTrivialUnswitch: Global variable that forces non-trivial 3591 // unswitching for testing and debugging. 3592 // NonTrivial: Parameter that enables non-trivial unswitching for this 3593 // invocation of the transform. But this should be allowed only 3594 // for targets without branch divergence. 3595 // 3596 // FIXME: If divergence analysis becomes available to a loop 3597 // transform, we should allow unswitching for non-trivial uniform 3598 // branches even on targets that have divergence. 3599 // https://bugs.llvm.org/show_bug.cgi?id=48819 3600 bool ContinueWithNonTrivial = 3601 EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F)); 3602 if (!ContinueWithNonTrivial) 3603 return false; 3604 3605 // Skip non-trivial unswitching for optsize functions. 3606 if (F->hasOptSize()) 3607 return false; 3608 3609 // Returns true if Loop L's loop nest is cold, i.e. if the headers of L, 3610 // of the loops L is nested in, and of the loops nested in L are all cold. 3611 auto IsLoopNestCold = [&](const Loop *L) { 3612 // Check L and all of its parent loops. 3613 auto *Parent = L; 3614 while (Parent) { 3615 if (!PSI->isColdBlock(Parent->getHeader(), BFI)) 3616 return false; 3617 Parent = Parent->getParentLoop(); 3618 } 3619 // Next check all loops nested within L. 3620 SmallVector<const Loop *, 4> Worklist; 3621 Worklist.insert(Worklist.end(), L->getSubLoops().begin(), 3622 L->getSubLoops().end()); 3623 while (!Worklist.empty()) { 3624 auto *CurLoop = Worklist.pop_back_val(); 3625 if (!PSI->isColdBlock(CurLoop->getHeader(), BFI)) 3626 return false; 3627 Worklist.insert(Worklist.end(), CurLoop->getSubLoops().begin(), 3628 CurLoop->getSubLoops().end()); 3629 } 3630 return true; 3631 }; 3632 3633 // Skip cold loops in cold loop nests, as unswitching them brings little 3634 // benefit but increases the code size 3635 if (PSI && PSI->hasProfileSummary() && BFI && IsLoopNestCold(&L)) { 3636 LLVM_DEBUG(dbgs() << " Skip cold loop: " << L << "\n"); 3637 return false; 3638 } 3639 3640 // Perform legality checks. 3641 if (!isSafeForNoNTrivialUnswitching(L, LI)) 3642 return false; 3643 3644 // For non-trivial unswitching, because it often creates new loops, we rely on 3645 // the pass manager to iterate on the loops rather than trying to immediately 3646 // reach a fixed point. There is no substantial advantage to iterating 3647 // internally, and if any of the new loops are simplified enough to contain 3648 // trivial unswitching we want to prefer those. 3649 3650 // Try to unswitch the best invariant condition. We prefer this full unswitch to 3651 // a partial unswitch when possible below the threshold. 3652 if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater)) 3653 return true; 3654 3655 // No other opportunities to unswitch. 3656 return false; 3657 } 3658 3659 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, 3660 LoopStandardAnalysisResults &AR, 3661 LPMUpdater &U) { 3662 Function &F = *L.getHeader()->getParent(); 3663 (void)F; 3664 ProfileSummaryInfo *PSI = nullptr; 3665 if (auto OuterProxy = 3666 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR) 3667 .getCachedResult<ModuleAnalysisManagerFunctionProxy>(F)) 3668 PSI = OuterProxy->getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); 3669 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L 3670 << "\n"); 3671 3672 std::optional<MemorySSAUpdater> MSSAU; 3673 if (AR.MSSA) { 3674 MSSAU = MemorySSAUpdater(AR.MSSA); 3675 if (VerifyMemorySSA) 3676 AR.MSSA->verifyMemorySSA(); 3677 } 3678 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial, 3679 &AR.SE, MSSAU ? &*MSSAU : nullptr, PSI, AR.BFI, U)) 3680 return PreservedAnalyses::all(); 3681 3682 if (AR.MSSA && VerifyMemorySSA) 3683 AR.MSSA->verifyMemorySSA(); 3684 3685 // Historically this pass has had issues with the dominator tree so verify it 3686 // in asserts builds. 3687 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast)); 3688 3689 auto PA = getLoopPassPreservedAnalyses(); 3690 if (AR.MSSA) 3691 PA.preserve<MemorySSAAnalysis>(); 3692 return PA; 3693 } 3694 3695 void SimpleLoopUnswitchPass::printPipeline( 3696 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 3697 static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline( 3698 OS, MapClassName2PassName); 3699 3700 OS << '<'; 3701 OS << (NonTrivial ? "" : "no-") << "nontrivial;"; 3702 OS << (Trivial ? "" : "no-") << "trivial"; 3703 OS << '>'; 3704 } 3705