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/Module.h" 45 #include "llvm/IR/PatternMatch.h" 46 #include "llvm/IR/ProfDataUtils.h" 47 #include "llvm/IR/Use.h" 48 #include "llvm/IR/Value.h" 49 #include "llvm/Support/Casting.h" 50 #include "llvm/Support/CommandLine.h" 51 #include "llvm/Support/Debug.h" 52 #include "llvm/Support/ErrorHandling.h" 53 #include "llvm/Support/GenericDomTree.h" 54 #include "llvm/Support/InstructionCost.h" 55 #include "llvm/Support/raw_ostream.h" 56 #include "llvm/Transforms/Scalar/LoopPassManager.h" 57 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 58 #include "llvm/Transforms/Utils/Cloning.h" 59 #include "llvm/Transforms/Utils/Local.h" 60 #include "llvm/Transforms/Utils/LoopUtils.h" 61 #include "llvm/Transforms/Utils/ValueMapper.h" 62 #include <algorithm> 63 #include <cassert> 64 #include <iterator> 65 #include <numeric> 66 #include <optional> 67 #include <utility> 68 69 #define DEBUG_TYPE "simple-loop-unswitch" 70 71 using namespace llvm; 72 using namespace llvm::PatternMatch; 73 74 STATISTIC(NumBranches, "Number of branches unswitched"); 75 STATISTIC(NumSwitches, "Number of switches unswitched"); 76 STATISTIC(NumSelects, "Number of selects turned into branches for unswitching"); 77 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching"); 78 STATISTIC(NumTrivial, "Number of unswitches that are trivial"); 79 STATISTIC( 80 NumCostMultiplierSkipped, 81 "Number of unswitch candidates that had their cost multiplier skipped"); 82 STATISTIC(NumInvariantConditionsInjected, 83 "Number of invariant conditions injected and unswitched"); 84 85 static cl::opt<bool> EnableNonTrivialUnswitch( 86 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden, 87 cl::desc("Forcibly enables non-trivial loop unswitching rather than " 88 "following the configuration passed into the pass.")); 89 90 static cl::opt<int> 91 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, 92 cl::desc("The cost threshold for unswitching a loop.")); 93 94 static cl::opt<bool> EnableUnswitchCostMultiplier( 95 "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden, 96 cl::desc("Enable unswitch cost multiplier that prohibits exponential " 97 "explosion in nontrivial unswitch.")); 98 static cl::opt<int> UnswitchSiblingsToplevelDiv( 99 "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden, 100 cl::desc("Toplevel siblings divisor for cost multiplier.")); 101 static cl::opt<int> UnswitchNumInitialUnscaledCandidates( 102 "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden, 103 cl::desc("Number of unswitch candidates that are ignored when calculating " 104 "cost multiplier.")); 105 static cl::opt<bool> UnswitchGuards( 106 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden, 107 cl::desc("If enabled, simple loop unswitching will also consider " 108 "llvm.experimental.guard intrinsics as unswitch candidates.")); 109 static cl::opt<bool> DropNonTrivialImplicitNullChecks( 110 "simple-loop-unswitch-drop-non-trivial-implicit-null-checks", 111 cl::init(false), cl::Hidden, 112 cl::desc("If enabled, drop make.implicit metadata in unswitched implicit " 113 "null checks to save time analyzing if we can keep it.")); 114 static cl::opt<unsigned> 115 MSSAThreshold("simple-loop-unswitch-memoryssa-threshold", 116 cl::desc("Max number of memory uses to explore during " 117 "partial unswitching analysis"), 118 cl::init(100), cl::Hidden); 119 static cl::opt<bool> FreezeLoopUnswitchCond( 120 "freeze-loop-unswitch-cond", cl::init(true), cl::Hidden, 121 cl::desc("If enabled, the freeze instruction will be added to condition " 122 "of loop unswitch to prevent miscompilation.")); 123 124 static cl::opt<bool> InjectInvariantConditions( 125 "simple-loop-unswitch-inject-invariant-conditions", cl::Hidden, 126 cl::desc("Whether we should inject new invariants and unswitch them to " 127 "eliminate some existing (non-invariant) conditions."), 128 cl::init(true)); 129 130 static cl::opt<unsigned> InjectInvariantConditionHotnesThreshold( 131 "simple-loop-unswitch-inject-invariant-condition-hotness-threshold", 132 cl::Hidden, cl::desc("Only try to inject loop invariant conditions and " 133 "unswitch on them to eliminate branches that are " 134 "not-taken 1/<this option> times or less."), 135 cl::init(16)); 136 137 AnalysisKey ShouldRunExtraSimpleLoopUnswitch::Key; 138 namespace { 139 struct CompareDesc { 140 BranchInst *Term; 141 Value *Invariant; 142 BasicBlock *InLoopSucc; 143 144 CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc) 145 : Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {} 146 }; 147 148 struct InjectedInvariant { 149 ICmpInst::Predicate Pred; 150 Value *LHS; 151 Value *RHS; 152 BasicBlock *InLoopSucc; 153 154 InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS, 155 BasicBlock *InLoopSucc) 156 : Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {} 157 }; 158 159 struct NonTrivialUnswitchCandidate { 160 Instruction *TI = nullptr; 161 TinyPtrVector<Value *> Invariants; 162 std::optional<InstructionCost> Cost; 163 std::optional<InjectedInvariant> PendingInjection; 164 NonTrivialUnswitchCandidate( 165 Instruction *TI, ArrayRef<Value *> Invariants, 166 std::optional<InstructionCost> Cost = std::nullopt, 167 std::optional<InjectedInvariant> PendingInjection = std::nullopt) 168 : TI(TI), Invariants(Invariants), Cost(Cost), 169 PendingInjection(PendingInjection) {}; 170 171 bool hasPendingInjection() const { return PendingInjection.has_value(); } 172 }; 173 } // end anonymous namespace. 174 175 // Helper to skip (select x, true, false), which matches both a logical AND and 176 // OR and can confuse code that tries to determine if \p Cond is either a 177 // logical AND or OR but not both. 178 static Value *skipTrivialSelect(Value *Cond) { 179 Value *CondNext; 180 while (match(Cond, m_Select(m_Value(CondNext), m_One(), m_Zero()))) 181 Cond = CondNext; 182 return Cond; 183 } 184 185 /// Collect all of the loop invariant input values transitively used by the 186 /// homogeneous instruction graph from a given root. 187 /// 188 /// This essentially walks from a root recursively through loop variant operands 189 /// which have perform the same logical operation (AND or OR) and finds all 190 /// inputs which are loop invariant. For some operations these can be 191 /// re-associated and unswitched out of the loop entirely. 192 static TinyPtrVector<Value *> 193 collectHomogenousInstGraphLoopInvariants(const Loop &L, Instruction &Root, 194 const LoopInfo &LI) { 195 assert(!L.isLoopInvariant(&Root) && 196 "Only need to walk the graph if root itself is not invariant."); 197 TinyPtrVector<Value *> Invariants; 198 199 bool IsRootAnd = match(&Root, m_LogicalAnd()); 200 bool IsRootOr = match(&Root, m_LogicalOr()); 201 202 // Build a worklist and recurse through operators collecting invariants. 203 SmallVector<Instruction *, 4> Worklist; 204 SmallPtrSet<Instruction *, 8> Visited; 205 Worklist.push_back(&Root); 206 Visited.insert(&Root); 207 do { 208 Instruction &I = *Worklist.pop_back_val(); 209 for (Value *OpV : I.operand_values()) { 210 // Skip constants as unswitching isn't interesting for them. 211 if (isa<Constant>(OpV)) 212 continue; 213 214 // Add it to our result if loop invariant. 215 if (L.isLoopInvariant(OpV)) { 216 Invariants.push_back(OpV); 217 continue; 218 } 219 220 // If not an instruction with the same opcode, nothing we can do. 221 Instruction *OpI = dyn_cast<Instruction>(skipTrivialSelect(OpV)); 222 223 if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) || 224 (IsRootOr && match(OpI, m_LogicalOr())))) { 225 // Visit this operand. 226 if (Visited.insert(OpI).second) 227 Worklist.push_back(OpI); 228 } 229 } 230 } while (!Worklist.empty()); 231 232 return Invariants; 233 } 234 235 static void replaceLoopInvariantUses(const Loop &L, Value *Invariant, 236 Constant &Replacement) { 237 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?"); 238 239 // Replace uses of LIC in the loop with the given constant. 240 // We use make_early_inc_range as set invalidates the iterator. 241 for (Use &U : llvm::make_early_inc_range(Invariant->uses())) { 242 Instruction *UserI = dyn_cast<Instruction>(U.getUser()); 243 244 // Replace this use within the loop body. 245 if (UserI && L.contains(UserI)) 246 U.set(&Replacement); 247 } 248 } 249 250 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial 251 /// incoming values along this edge. 252 static bool areLoopExitPHIsLoopInvariant(const Loop &L, 253 const BasicBlock &ExitingBB, 254 const BasicBlock &ExitBB) { 255 for (const Instruction &I : ExitBB) { 256 auto *PN = dyn_cast<PHINode>(&I); 257 if (!PN) 258 // No more PHIs to check. 259 return true; 260 261 // If the incoming value for this edge isn't loop invariant the unswitch 262 // won't be trivial. 263 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))) 264 return false; 265 } 266 llvm_unreachable("Basic blocks should never be empty!"); 267 } 268 269 /// Copy a set of loop invariant values \p ToDuplicate and insert them at the 270 /// end of \p BB and conditionally branch on the copied condition. We only 271 /// branch on a single value. 272 static void buildPartialUnswitchConditionalBranch( 273 BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction, 274 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze, 275 const Instruction *I, AssumptionCache *AC, const DominatorTree &DT) { 276 IRBuilder<> IRB(&BB); 277 278 SmallVector<Value *> FrozenInvariants; 279 for (Value *Inv : Invariants) { 280 if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(Inv, AC, I, &DT)) 281 Inv = IRB.CreateFreeze(Inv, Inv->getName() + ".fr"); 282 FrozenInvariants.push_back(Inv); 283 } 284 285 Value *Cond = Direction ? IRB.CreateOr(FrozenInvariants) 286 : IRB.CreateAnd(FrozenInvariants); 287 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, 288 Direction ? &NormalSucc : &UnswitchedSucc); 289 } 290 291 /// Copy a set of loop invariant values, and conditionally branch on them. 292 static void buildPartialInvariantUnswitchConditionalBranch( 293 BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction, 294 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L, 295 MemorySSAUpdater *MSSAU) { 296 ValueToValueMapTy VMap; 297 for (auto *Val : reverse(ToDuplicate)) { 298 Instruction *Inst = cast<Instruction>(Val); 299 Instruction *NewInst = Inst->clone(); 300 NewInst->insertInto(&BB, BB.end()); 301 RemapInstruction(NewInst, VMap, 302 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 303 VMap[Val] = NewInst; 304 305 if (!MSSAU) 306 continue; 307 308 MemorySSA *MSSA = MSSAU->getMemorySSA(); 309 if (auto *MemUse = 310 dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) { 311 auto *DefiningAccess = MemUse->getDefiningAccess(); 312 // Get the first defining access before the loop. 313 while (L.contains(DefiningAccess->getBlock())) { 314 // If the defining access is a MemoryPhi, get the incoming 315 // value for the pre-header as defining access. 316 if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess)) 317 DefiningAccess = 318 MemPhi->getIncomingValueForBlock(L.getLoopPreheader()); 319 else 320 DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess(); 321 } 322 MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess, 323 NewInst->getParent(), 324 MemorySSA::BeforeTerminator); 325 } 326 } 327 328 IRBuilder<> IRB(&BB); 329 Value *Cond = VMap[ToDuplicate[0]]; 330 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, 331 Direction ? &NormalSucc : &UnswitchedSucc); 332 } 333 334 /// Rewrite the PHI nodes in an unswitched loop exit basic block. 335 /// 336 /// Requires that the loop exit and unswitched basic block are the same, and 337 /// that the exiting block was a unique predecessor of that block. Rewrites the 338 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial 339 /// PHI nodes from the old preheader that now contains the unswitched 340 /// terminator. 341 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, 342 BasicBlock &OldExitingBB, 343 BasicBlock &OldPH) { 344 for (PHINode &PN : UnswitchedBB.phis()) { 345 // When the loop exit is directly unswitched we just need to update the 346 // incoming basic block. We loop to handle weird cases with repeated 347 // incoming blocks, but expect to typically only have one operand here. 348 for (auto i : seq<int>(0, PN.getNumOperands())) { 349 assert(PN.getIncomingBlock(i) == &OldExitingBB && 350 "Found incoming block different from unique predecessor!"); 351 PN.setIncomingBlock(i, &OldPH); 352 } 353 } 354 } 355 356 /// Rewrite the PHI nodes in the loop exit basic block and the split off 357 /// unswitched block. 358 /// 359 /// Because the exit block remains an exit from the loop, this rewrites the 360 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI 361 /// nodes into the unswitched basic block to select between the value in the 362 /// old preheader and the loop exit. 363 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, 364 BasicBlock &UnswitchedBB, 365 BasicBlock &OldExitingBB, 366 BasicBlock &OldPH, 367 bool FullUnswitch) { 368 assert(&ExitBB != &UnswitchedBB && 369 "Must have different loop exit and unswitched blocks!"); 370 BasicBlock::iterator InsertPt = UnswitchedBB.begin(); 371 for (PHINode &PN : ExitBB.phis()) { 372 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2, 373 PN.getName() + ".split"); 374 NewPN->insertBefore(InsertPt); 375 376 // Walk backwards over the old PHI node's inputs to minimize the cost of 377 // removing each one. We have to do this weird loop manually so that we 378 // create the same number of new incoming edges in the new PHI as we expect 379 // each case-based edge to be included in the unswitched switch in some 380 // cases. 381 // FIXME: This is really, really gross. It would be much cleaner if LLVM 382 // allowed us to create a single entry for a predecessor block without 383 // having separate entries for each "edge" even though these edges are 384 // required to produce identical results. 385 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) { 386 if (PN.getIncomingBlock(i) != &OldExitingBB) 387 continue; 388 389 Value *Incoming = PN.getIncomingValue(i); 390 if (FullUnswitch) 391 // No more edge from the old exiting block to the exit block. 392 PN.removeIncomingValue(i); 393 394 NewPN->addIncoming(Incoming, &OldPH); 395 } 396 397 // Now replace the old PHI with the new one and wire the old one in as an 398 // input to the new one. 399 PN.replaceAllUsesWith(NewPN); 400 NewPN->addIncoming(&PN, &ExitBB); 401 } 402 } 403 404 /// Hoist the current loop up to the innermost loop containing a remaining exit. 405 /// 406 /// Because we've removed an exit from the loop, we may have changed the set of 407 /// loops reachable and need to move the current loop up the loop nest or even 408 /// to an entirely separate nest. 409 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader, 410 DominatorTree &DT, LoopInfo &LI, 411 MemorySSAUpdater *MSSAU, ScalarEvolution *SE) { 412 // If the loop is already at the top level, we can't hoist it anywhere. 413 Loop *OldParentL = L.getParentLoop(); 414 if (!OldParentL) 415 return; 416 417 SmallVector<BasicBlock *, 4> Exits; 418 L.getExitBlocks(Exits); 419 Loop *NewParentL = nullptr; 420 for (auto *ExitBB : Exits) 421 if (Loop *ExitL = LI.getLoopFor(ExitBB)) 422 if (!NewParentL || NewParentL->contains(ExitL)) 423 NewParentL = ExitL; 424 425 if (NewParentL == OldParentL) 426 return; 427 428 // The new parent loop (if different) should always contain the old one. 429 if (NewParentL) 430 assert(NewParentL->contains(OldParentL) && 431 "Can only hoist this loop up the nest!"); 432 433 // The preheader will need to move with the body of this loop. However, 434 // because it isn't in this loop we also need to update the primary loop map. 435 assert(OldParentL == LI.getLoopFor(&Preheader) && 436 "Parent loop of this loop should contain this loop's preheader!"); 437 LI.changeLoopFor(&Preheader, NewParentL); 438 439 // Remove this loop from its old parent. 440 OldParentL->removeChildLoop(&L); 441 442 // Add the loop either to the new parent or as a top-level loop. 443 if (NewParentL) 444 NewParentL->addChildLoop(&L); 445 else 446 LI.addTopLevelLoop(&L); 447 448 // Remove this loops blocks from the old parent and every other loop up the 449 // nest until reaching the new parent. Also update all of these 450 // no-longer-containing loops to reflect the nesting change. 451 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL; 452 OldContainingL = OldContainingL->getParentLoop()) { 453 llvm::erase_if(OldContainingL->getBlocksVector(), 454 [&](const BasicBlock *BB) { 455 return BB == &Preheader || L.contains(BB); 456 }); 457 458 OldContainingL->getBlocksSet().erase(&Preheader); 459 for (BasicBlock *BB : L.blocks()) 460 OldContainingL->getBlocksSet().erase(BB); 461 462 // Because we just hoisted a loop out of this one, we have essentially 463 // created new exit paths from it. That means we need to form LCSSA PHI 464 // nodes for values used in the no-longer-nested loop. 465 formLCSSA(*OldContainingL, DT, &LI, SE); 466 467 // We shouldn't need to form dedicated exits because the exit introduced 468 // here is the (just split by unswitching) preheader. However, after trivial 469 // unswitching it is possible to get new non-dedicated exits out of parent 470 // loop so let's conservatively form dedicated exit blocks and figure out 471 // if we can optimize later. 472 formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU, 473 /*PreserveLCSSA*/ true); 474 } 475 } 476 477 // Return the top-most loop containing ExitBB and having ExitBB as exiting block 478 // or the loop containing ExitBB, if there is no parent loop containing ExitBB 479 // as exiting block. 480 static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB, 481 const LoopInfo &LI) { 482 Loop *TopMost = LI.getLoopFor(ExitBB); 483 Loop *Current = TopMost; 484 while (Current) { 485 if (Current->isLoopExiting(ExitBB)) 486 TopMost = Current; 487 Current = Current->getParentLoop(); 488 } 489 return TopMost; 490 } 491 492 /// Unswitch a trivial branch if the condition is loop invariant. 493 /// 494 /// This routine should only be called when loop code leading to the branch has 495 /// been validated as trivial (no side effects). This routine checks if the 496 /// condition is invariant and one of the successors is a loop exit. This 497 /// allows us to unswitch without duplicating the loop, making it trivial. 498 /// 499 /// If this routine fails to unswitch the branch it returns false. 500 /// 501 /// If the branch can be unswitched, this routine splits the preheader and 502 /// hoists the branch above that split. Preserves loop simplified form 503 /// (splitting the exit block as necessary). It simplifies the branch within 504 /// the loop to an unconditional branch but doesn't remove it entirely. Further 505 /// cleanup can be done with some simplifycfg like pass. 506 /// 507 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 508 /// invalidated by this. 509 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, 510 LoopInfo &LI, ScalarEvolution *SE, 511 MemorySSAUpdater *MSSAU) { 512 assert(BI.isConditional() && "Can only unswitch a conditional branch!"); 513 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n"); 514 515 // The loop invariant values that we want to unswitch. 516 TinyPtrVector<Value *> Invariants; 517 518 // When true, we're fully unswitching the branch rather than just unswitching 519 // some input conditions to the branch. 520 bool FullUnswitch = false; 521 522 Value *Cond = skipTrivialSelect(BI.getCondition()); 523 if (L.isLoopInvariant(Cond)) { 524 Invariants.push_back(Cond); 525 FullUnswitch = true; 526 } else { 527 if (auto *CondInst = dyn_cast<Instruction>(Cond)) 528 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI); 529 if (Invariants.empty()) { 530 LLVM_DEBUG(dbgs() << " Couldn't find invariant inputs!\n"); 531 return false; 532 } 533 } 534 535 // Check that one of the branch's successors exits, and which one. 536 bool ExitDirection = true; 537 int LoopExitSuccIdx = 0; 538 auto *LoopExitBB = BI.getSuccessor(0); 539 if (L.contains(LoopExitBB)) { 540 ExitDirection = false; 541 LoopExitSuccIdx = 1; 542 LoopExitBB = BI.getSuccessor(1); 543 if (L.contains(LoopExitBB)) { 544 LLVM_DEBUG(dbgs() << " Branch doesn't exit the loop!\n"); 545 return false; 546 } 547 } 548 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); 549 auto *ParentBB = BI.getParent(); 550 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) { 551 LLVM_DEBUG(dbgs() << " Loop exit PHI's aren't loop-invariant!\n"); 552 return false; 553 } 554 555 // When unswitching only part of the branch's condition, we need the exit 556 // block to be reached directly from the partially unswitched input. This can 557 // be done when the exit block is along the true edge and the branch condition 558 // is a graph of `or` operations, or the exit block is along the false edge 559 // and the condition is a graph of `and` operations. 560 if (!FullUnswitch) { 561 if (ExitDirection ? !match(Cond, m_LogicalOr()) 562 : !match(Cond, m_LogicalAnd())) { 563 LLVM_DEBUG(dbgs() << " Branch condition is in improper form for " 564 "non-full unswitch!\n"); 565 return false; 566 } 567 } 568 569 LLVM_DEBUG({ 570 dbgs() << " unswitching trivial invariant conditions for: " << BI 571 << "\n"; 572 for (Value *Invariant : Invariants) { 573 dbgs() << " " << *Invariant << " == true"; 574 if (Invariant != Invariants.back()) 575 dbgs() << " ||"; 576 dbgs() << "\n"; 577 } 578 }); 579 580 // If we have scalar evolutions, we need to invalidate them including this 581 // loop, the loop containing the exit block and the topmost parent loop 582 // exiting via LoopExitBB. 583 if (SE) { 584 if (const Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI)) 585 SE->forgetLoop(ExitL); 586 else 587 // Forget the entire nest as this exits the entire nest. 588 SE->forgetTopmostLoop(&L); 589 SE->forgetBlockAndLoopDispositions(); 590 } 591 592 if (MSSAU && VerifyMemorySSA) 593 MSSAU->getMemorySSA()->verifyMemorySSA(); 594 595 // Split the preheader, so that we know that there is a safe place to insert 596 // the conditional branch. We will change the preheader to have a conditional 597 // branch on LoopCond. 598 BasicBlock *OldPH = L.getLoopPreheader(); 599 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); 600 601 // Now that we have a place to insert the conditional branch, create a place 602 // to branch to: this is the exit block out of the loop that we are 603 // unswitching. We need to split this if there are other loop predecessors. 604 // Because the loop is in simplified form, *any* other predecessor is enough. 605 BasicBlock *UnswitchedBB; 606 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) { 607 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() && 608 "A branch's parent isn't a predecessor!"); 609 UnswitchedBB = LoopExitBB; 610 } else { 611 UnswitchedBB = 612 SplitBlock(LoopExitBB, LoopExitBB->begin(), &DT, &LI, MSSAU, "", false); 613 } 614 615 if (MSSAU && VerifyMemorySSA) 616 MSSAU->getMemorySSA()->verifyMemorySSA(); 617 618 // Actually move the invariant uses into the unswitched position. If possible, 619 // we do this by moving the instructions, but when doing partial unswitching 620 // we do it by building a new merge of the values in the unswitched position. 621 OldPH->getTerminator()->eraseFromParent(); 622 if (FullUnswitch) { 623 // If fully unswitching, we can use the existing branch instruction. 624 // Splice it into the old PH to gate reaching the new preheader and re-point 625 // its successors. 626 BI.moveBefore(*OldPH, OldPH->end()); 627 BI.setCondition(Cond); 628 if (MSSAU) { 629 // Temporarily clone the terminator, to make MSSA update cheaper by 630 // separating "insert edge" updates from "remove edge" ones. 631 BI.clone()->insertInto(ParentBB, ParentBB->end()); 632 } else { 633 // Create a new unconditional branch that will continue the loop as a new 634 // terminator. 635 Instruction *NewBI = BranchInst::Create(ContinueBB, ParentBB); 636 NewBI->setDebugLoc(BI.getDebugLoc()); 637 } 638 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); 639 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); 640 } else { 641 // Only unswitching a subset of inputs to the condition, so we will need to 642 // build a new branch that merges the invariant inputs. 643 if (ExitDirection) 644 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) && 645 "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the " 646 "condition!"); 647 else 648 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) && 649 "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the" 650 " condition!"); 651 buildPartialUnswitchConditionalBranch( 652 *OldPH, Invariants, ExitDirection, *UnswitchedBB, *NewPH, 653 FreezeLoopUnswitchCond, OldPH->getTerminator(), nullptr, DT); 654 } 655 656 // Update the dominator tree with the added edge. 657 DT.insertEdge(OldPH, UnswitchedBB); 658 659 // After the dominator tree was updated with the added edge, update MemorySSA 660 // if available. 661 if (MSSAU) { 662 SmallVector<CFGUpdate, 1> Updates; 663 Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB}); 664 MSSAU->applyInsertUpdates(Updates, DT); 665 } 666 667 // Finish updating dominator tree and memory ssa for full unswitch. 668 if (FullUnswitch) { 669 if (MSSAU) { 670 Instruction *Term = ParentBB->getTerminator(); 671 // Remove the cloned branch instruction and create unconditional branch 672 // now. 673 Instruction *NewBI = BranchInst::Create(ContinueBB, ParentBB); 674 NewBI->setDebugLoc(Term->getDebugLoc()); 675 Term->eraseFromParent(); 676 MSSAU->removeEdge(ParentBB, LoopExitBB); 677 } 678 DT.deleteEdge(ParentBB, LoopExitBB); 679 } 680 681 if (MSSAU && VerifyMemorySSA) 682 MSSAU->getMemorySSA()->verifyMemorySSA(); 683 684 // Rewrite the relevant PHI nodes. 685 if (UnswitchedBB == LoopExitBB) 686 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); 687 else 688 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, 689 *ParentBB, *OldPH, FullUnswitch); 690 691 // The constant we can replace all of our invariants with inside the loop 692 // body. If any of the invariants have a value other than this the loop won't 693 // be entered. 694 ConstantInt *Replacement = ExitDirection 695 ? ConstantInt::getFalse(BI.getContext()) 696 : ConstantInt::getTrue(BI.getContext()); 697 698 // Since this is an i1 condition we can also trivially replace uses of it 699 // within the loop with a constant. 700 for (Value *Invariant : Invariants) 701 replaceLoopInvariantUses(L, Invariant, *Replacement); 702 703 // If this was full unswitching, we may have changed the nesting relationship 704 // for this loop so hoist it to its correct parent if needed. 705 if (FullUnswitch) 706 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE); 707 708 if (MSSAU && VerifyMemorySSA) 709 MSSAU->getMemorySSA()->verifyMemorySSA(); 710 711 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n"); 712 ++NumTrivial; 713 ++NumBranches; 714 return true; 715 } 716 717 /// Unswitch a trivial switch if the condition is loop invariant. 718 /// 719 /// This routine should only be called when loop code leading to the switch has 720 /// been validated as trivial (no side effects). This routine checks if the 721 /// condition is invariant and that at least one of the successors is a loop 722 /// exit. This allows us to unswitch without duplicating the loop, making it 723 /// trivial. 724 /// 725 /// If this routine fails to unswitch the switch it returns false. 726 /// 727 /// If the switch can be unswitched, this routine splits the preheader and 728 /// copies the switch above that split. If the default case is one of the 729 /// exiting cases, it copies the non-exiting cases and points them at the new 730 /// preheader. If the default case is not exiting, it copies the exiting cases 731 /// and points the default at the preheader. It preserves loop simplified form 732 /// (splitting the exit blocks as necessary). It simplifies the switch within 733 /// the loop by removing now-dead cases. If the default case is one of those 734 /// unswitched, it replaces its destination with a new basic block containing 735 /// only unreachable. Such basic blocks, while technically loop exits, are not 736 /// considered for unswitching so this is a stable transform and the same 737 /// switch will not be revisited. If after unswitching there is only a single 738 /// in-loop successor, the switch is further simplified to an unconditional 739 /// branch. Still more cleanup can be done with some simplifycfg like pass. 740 /// 741 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 742 /// invalidated by this. 743 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, 744 LoopInfo &LI, ScalarEvolution *SE, 745 MemorySSAUpdater *MSSAU) { 746 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n"); 747 Value *LoopCond = SI.getCondition(); 748 749 // If this isn't switching on an invariant condition, we can't unswitch it. 750 if (!L.isLoopInvariant(LoopCond)) 751 return false; 752 753 auto *ParentBB = SI.getParent(); 754 755 // The same check must be used both for the default and the exit cases. We 756 // should never leave edges from the switch instruction to a basic block that 757 // we are unswitching, hence the condition used to determine the default case 758 // needs to also be used to populate ExitCaseIndices, which is then used to 759 // remove cases from the switch. 760 auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) { 761 // BBToCheck is not an exit block if it is inside loop L. 762 if (L.contains(&BBToCheck)) 763 return false; 764 // BBToCheck is not trivial to unswitch if its phis aren't loop invariant. 765 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck)) 766 return false; 767 // We do not unswitch a block that only has an unreachable statement, as 768 // it's possible this is a previously unswitched block. Only unswitch if 769 // either the terminator is not unreachable, or, if it is, it's not the only 770 // instruction in the block. 771 auto *TI = BBToCheck.getTerminator(); 772 bool isUnreachable = isa<UnreachableInst>(TI); 773 return !isUnreachable || 774 (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI)); 775 }; 776 777 SmallVector<int, 4> ExitCaseIndices; 778 for (auto Case : SI.cases()) 779 if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor())) 780 ExitCaseIndices.push_back(Case.getCaseIndex()); 781 BasicBlock *DefaultExitBB = nullptr; 782 SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight = 783 SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0); 784 if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) { 785 DefaultExitBB = SI.getDefaultDest(); 786 } else if (ExitCaseIndices.empty()) 787 return false; 788 789 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n"); 790 791 if (MSSAU && VerifyMemorySSA) 792 MSSAU->getMemorySSA()->verifyMemorySSA(); 793 794 // We may need to invalidate SCEVs for the outermost loop reached by any of 795 // the exits. 796 Loop *OuterL = &L; 797 798 if (DefaultExitBB) { 799 // Check the loop containing this exit. 800 Loop *ExitL = getTopMostExitingLoop(DefaultExitBB, LI); 801 if (!ExitL || ExitL->contains(OuterL)) 802 OuterL = ExitL; 803 } 804 for (unsigned Index : ExitCaseIndices) { 805 auto CaseI = SI.case_begin() + Index; 806 // Compute the outer loop from this exit. 807 Loop *ExitL = getTopMostExitingLoop(CaseI->getCaseSuccessor(), LI); 808 if (!ExitL || ExitL->contains(OuterL)) 809 OuterL = ExitL; 810 } 811 812 if (SE) { 813 if (OuterL) 814 SE->forgetLoop(OuterL); 815 else 816 SE->forgetTopmostLoop(&L); 817 } 818 819 if (DefaultExitBB) { 820 // Clear out the default destination temporarily to allow accurate 821 // predecessor lists to be examined below. 822 SI.setDefaultDest(nullptr); 823 } 824 825 // Store the exit cases into a separate data structure and remove them from 826 // the switch. 827 SmallVector<std::tuple<ConstantInt *, BasicBlock *, 828 SwitchInstProfUpdateWrapper::CaseWeightOpt>, 829 4> ExitCases; 830 ExitCases.reserve(ExitCaseIndices.size()); 831 SwitchInstProfUpdateWrapper SIW(SI); 832 // We walk the case indices backwards so that we remove the last case first 833 // and don't disrupt the earlier indices. 834 for (unsigned Index : reverse(ExitCaseIndices)) { 835 auto CaseI = SI.case_begin() + Index; 836 // Save the value of this case. 837 auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex()); 838 ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W); 839 // Delete the unswitched cases. 840 SIW.removeCase(CaseI); 841 } 842 843 // Check if after this all of the remaining cases point at the same 844 // successor. 845 BasicBlock *CommonSuccBB = nullptr; 846 if (SI.getNumCases() > 0 && 847 all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) { 848 return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor(); 849 })) 850 CommonSuccBB = SI.case_begin()->getCaseSuccessor(); 851 if (!DefaultExitBB) { 852 // If we're not unswitching the default, we need it to match any cases to 853 // have a common successor or if we have no cases it is the common 854 // successor. 855 if (SI.getNumCases() == 0) 856 CommonSuccBB = SI.getDefaultDest(); 857 else if (SI.getDefaultDest() != CommonSuccBB) 858 CommonSuccBB = nullptr; 859 } 860 861 // Split the preheader, so that we know that there is a safe place to insert 862 // the switch. 863 BasicBlock *OldPH = L.getLoopPreheader(); 864 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); 865 OldPH->getTerminator()->eraseFromParent(); 866 867 // Now add the unswitched switch. This new switch instruction inherits the 868 // debug location of the old switch, because it semantically replace the old 869 // one. 870 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); 871 NewSI->setDebugLoc(SIW->getDebugLoc()); 872 SwitchInstProfUpdateWrapper NewSIW(*NewSI); 873 874 // Rewrite the IR for the unswitched basic blocks. This requires two steps. 875 // First, we split any exit blocks with remaining in-loop predecessors. Then 876 // we update the PHIs in one of two ways depending on if there was a split. 877 // We walk in reverse so that we split in the same order as the cases 878 // appeared. This is purely for convenience of reading the resulting IR, but 879 // it doesn't cost anything really. 880 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; 881 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; 882 // Handle the default exit if necessary. 883 // FIXME: It'd be great if we could merge this with the loop below but LLVM's 884 // ranges aren't quite powerful enough yet. 885 if (DefaultExitBB) { 886 if (pred_empty(DefaultExitBB)) { 887 UnswitchedExitBBs.insert(DefaultExitBB); 888 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); 889 } else { 890 auto *SplitBB = 891 SplitBlock(DefaultExitBB, DefaultExitBB->begin(), &DT, &LI, MSSAU); 892 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB, 893 *ParentBB, *OldPH, 894 /*FullUnswitch*/ true); 895 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; 896 } 897 } 898 // Note that we must use a reference in the for loop so that we update the 899 // container. 900 for (auto &ExitCase : reverse(ExitCases)) { 901 // Grab a reference to the exit block in the pair so that we can update it. 902 BasicBlock *ExitBB = std::get<1>(ExitCase); 903 904 // If this case is the last edge into the exit block, we can simply reuse it 905 // as it will no longer be a loop exit. No mapping necessary. 906 if (pred_empty(ExitBB)) { 907 // Only rewrite once. 908 if (UnswitchedExitBBs.insert(ExitBB).second) 909 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); 910 continue; 911 } 912 913 // Otherwise we need to split the exit block so that we retain an exit 914 // block from the loop and a target for the unswitched condition. 915 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; 916 if (!SplitExitBB) { 917 // If this is the first time we see this, do the split and remember it. 918 SplitExitBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU); 919 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB, 920 *ParentBB, *OldPH, 921 /*FullUnswitch*/ true); 922 } 923 // Update the case pair to point to the split block. 924 std::get<1>(ExitCase) = SplitExitBB; 925 } 926 927 // Now add the unswitched cases. We do this in reverse order as we built them 928 // in reverse order. 929 for (auto &ExitCase : reverse(ExitCases)) { 930 ConstantInt *CaseVal = std::get<0>(ExitCase); 931 BasicBlock *UnswitchedBB = std::get<1>(ExitCase); 932 933 NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase)); 934 } 935 936 // If the default was unswitched, re-point it and add explicit cases for 937 // entering the loop. 938 if (DefaultExitBB) { 939 NewSIW->setDefaultDest(DefaultExitBB); 940 NewSIW.setSuccessorWeight(0, DefaultCaseWeight); 941 942 // We removed all the exit cases, so we just copy the cases to the 943 // unswitched switch. 944 for (const auto &Case : SI.cases()) 945 NewSIW.addCase(Case.getCaseValue(), NewPH, 946 SIW.getSuccessorWeight(Case.getSuccessorIndex())); 947 } else if (DefaultCaseWeight) { 948 // We have to set branch weight of the default case. 949 uint64_t SW = *DefaultCaseWeight; 950 for (const auto &Case : SI.cases()) { 951 auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex()); 952 assert(W && 953 "case weight must be defined as default case weight is defined"); 954 SW += *W; 955 } 956 NewSIW.setSuccessorWeight(0, SW); 957 } 958 959 // If we ended up with a common successor for every path through the switch 960 // after unswitching, rewrite it to an unconditional branch to make it easy 961 // to recognize. Otherwise we potentially have to recognize the default case 962 // pointing at unreachable and other complexity. 963 if (CommonSuccBB) { 964 BasicBlock *BB = SI.getParent(); 965 // We may have had multiple edges to this common successor block, so remove 966 // them as predecessors. We skip the first one, either the default or the 967 // actual first case. 968 bool SkippedFirst = DefaultExitBB == nullptr; 969 for (auto Case : SI.cases()) { 970 assert(Case.getCaseSuccessor() == CommonSuccBB && 971 "Non-common successor!"); 972 (void)Case; 973 if (!SkippedFirst) { 974 SkippedFirst = true; 975 continue; 976 } 977 CommonSuccBB->removePredecessor(BB, 978 /*KeepOneInputPHIs*/ true); 979 } 980 // Now nuke the switch and replace it with a direct branch. 981 Instruction *NewBI = BranchInst::Create(CommonSuccBB, BB); 982 NewBI->setDebugLoc(SIW->getDebugLoc()); 983 SIW.eraseFromParent(); 984 } else if (DefaultExitBB) { 985 assert(SI.getNumCases() > 0 && 986 "If we had no cases we'd have a common successor!"); 987 // Move the last case to the default successor. This is valid as if the 988 // default got unswitched it cannot be reached. This has the advantage of 989 // being simple and keeping the number of edges from this switch to 990 // successors the same, and avoiding any PHI update complexity. 991 auto LastCaseI = std::prev(SI.case_end()); 992 993 SI.setDefaultDest(LastCaseI->getCaseSuccessor()); 994 SIW.setSuccessorWeight( 995 0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex())); 996 SIW.removeCase(LastCaseI); 997 } 998 999 // Walk the unswitched exit blocks and the unswitched split blocks and update 1000 // the dominator tree based on the CFG edits. While we are walking unordered 1001 // containers here, the API for applyUpdates takes an unordered list of 1002 // updates and requires them to not contain duplicates. 1003 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 1004 for (auto *UnswitchedExitBB : UnswitchedExitBBs) { 1005 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB}); 1006 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB}); 1007 } 1008 for (auto SplitUnswitchedPair : SplitExitBBMap) { 1009 DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first}); 1010 DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second}); 1011 } 1012 1013 if (MSSAU) { 1014 MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true); 1015 if (VerifyMemorySSA) 1016 MSSAU->getMemorySSA()->verifyMemorySSA(); 1017 } else { 1018 DT.applyUpdates(DTUpdates); 1019 } 1020 1021 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 1022 1023 // We may have changed the nesting relationship for this loop so hoist it to 1024 // its correct parent if needed. 1025 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE); 1026 1027 if (MSSAU && VerifyMemorySSA) 1028 MSSAU->getMemorySSA()->verifyMemorySSA(); 1029 1030 ++NumTrivial; 1031 ++NumSwitches; 1032 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n"); 1033 return true; 1034 } 1035 1036 /// This routine scans the loop to find a branch or switch which occurs before 1037 /// any side effects occur. These can potentially be unswitched without 1038 /// duplicating the loop. If a branch or switch is successfully unswitched the 1039 /// scanning continues to see if subsequent branches or switches have become 1040 /// trivial. Once all trivial candidates have been unswitched, this routine 1041 /// returns. 1042 /// 1043 /// The return value indicates whether anything was unswitched (and therefore 1044 /// changed). 1045 /// 1046 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 1047 /// invalidated by this. 1048 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, 1049 LoopInfo &LI, ScalarEvolution *SE, 1050 MemorySSAUpdater *MSSAU) { 1051 bool Changed = false; 1052 1053 // If loop header has only one reachable successor we should keep looking for 1054 // trivial condition candidates in the successor as well. An alternative is 1055 // to constant fold conditions and merge successors into loop header (then we 1056 // only need to check header's terminator). The reason for not doing this in 1057 // LoopUnswitch pass is that it could potentially break LoopPassManager's 1058 // invariants. Folding dead branches could either eliminate the current loop 1059 // or make other loops unreachable. LCSSA form might also not be preserved 1060 // after deleting branches. The following code keeps traversing loop header's 1061 // successors until it finds the trivial condition candidate (condition that 1062 // is not a constant). Since unswitching generates branches with constant 1063 // conditions, this scenario could be very common in practice. 1064 BasicBlock *CurrentBB = L.getHeader(); 1065 SmallPtrSet<BasicBlock *, 8> Visited; 1066 Visited.insert(CurrentBB); 1067 do { 1068 // Check if there are any side-effecting instructions (e.g. stores, calls, 1069 // volatile loads) in the part of the loop that the code *would* execute 1070 // without unswitching. 1071 if (MSSAU) // Possible early exit with MSSA 1072 if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB)) 1073 if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end())) 1074 return Changed; 1075 if (llvm::any_of(*CurrentBB, 1076 [](Instruction &I) { return I.mayHaveSideEffects(); })) 1077 return Changed; 1078 1079 Instruction *CurrentTerm = CurrentBB->getTerminator(); 1080 1081 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) { 1082 // Don't bother trying to unswitch past a switch with a constant 1083 // condition. This should be removed prior to running this pass by 1084 // simplifycfg. 1085 if (isa<Constant>(SI->getCondition())) 1086 return Changed; 1087 1088 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU)) 1089 // Couldn't unswitch this one so we're done. 1090 return Changed; 1091 1092 // Mark that we managed to unswitch something. 1093 Changed = true; 1094 1095 // If unswitching turned the terminator into an unconditional branch then 1096 // we can continue. The unswitching logic specifically works to fold any 1097 // cases it can into an unconditional branch to make it easier to 1098 // recognize here. 1099 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); 1100 if (!BI || BI->isConditional()) 1101 return Changed; 1102 1103 CurrentBB = BI->getSuccessor(0); 1104 continue; 1105 } 1106 1107 auto *BI = dyn_cast<BranchInst>(CurrentTerm); 1108 if (!BI) 1109 // We do not understand other terminator instructions. 1110 return Changed; 1111 1112 // Don't bother trying to unswitch past an unconditional branch or a branch 1113 // with a constant value. These should be removed by simplifycfg prior to 1114 // running this pass. 1115 if (!BI->isConditional() || 1116 isa<Constant>(skipTrivialSelect(BI->getCondition()))) 1117 return Changed; 1118 1119 // Found a trivial condition candidate: non-foldable conditional branch. If 1120 // we fail to unswitch this, we can't do anything else that is trivial. 1121 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU)) 1122 return Changed; 1123 1124 // Mark that we managed to unswitch something. 1125 Changed = true; 1126 1127 // If we only unswitched some of the conditions feeding the branch, we won't 1128 // have collapsed it to a single successor. 1129 BI = cast<BranchInst>(CurrentBB->getTerminator()); 1130 if (BI->isConditional()) 1131 return Changed; 1132 1133 // Follow the newly unconditional branch into its successor. 1134 CurrentBB = BI->getSuccessor(0); 1135 1136 // When continuing, if we exit the loop or reach a previous visited block, 1137 // then we can not reach any trivial condition candidates (unfoldable 1138 // branch instructions or switch instructions) and no unswitch can happen. 1139 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second); 1140 1141 return Changed; 1142 } 1143 1144 /// Build the cloned blocks for an unswitched copy of the given loop. 1145 /// 1146 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and 1147 /// after the split block (`SplitBB`) that will be used to select between the 1148 /// cloned and original loop. 1149 /// 1150 /// This routine handles cloning all of the necessary loop blocks and exit 1151 /// blocks including rewriting their instructions and the relevant PHI nodes. 1152 /// Any loop blocks or exit blocks which are dominated by a different successor 1153 /// than the one for this clone of the loop blocks can be trivially skipped. We 1154 /// use the `DominatingSucc` map to determine whether a block satisfies that 1155 /// property with a simple map lookup. 1156 /// 1157 /// It also correctly creates the unconditional branch in the cloned 1158 /// unswitched parent block to only point at the unswitched successor. 1159 /// 1160 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit 1161 /// block splitting is correctly reflected in `LoopInfo`, essentially all of 1162 /// the cloned blocks (and their loops) are left without full `LoopInfo` 1163 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned 1164 /// blocks to them but doesn't create the cloned `DominatorTree` structure and 1165 /// instead the caller must recompute an accurate DT. It *does* correctly 1166 /// update the `AssumptionCache` provided in `AC`. 1167 static BasicBlock *buildClonedLoopBlocks( 1168 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, 1169 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB, 1170 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, 1171 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc, 1172 ValueToValueMapTy &VMap, 1173 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC, 1174 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU, 1175 ScalarEvolution *SE) { 1176 SmallVector<BasicBlock *, 4> NewBlocks; 1177 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size()); 1178 1179 // We will need to clone a bunch of blocks, wrap up the clone operation in 1180 // a helper. 1181 auto CloneBlock = [&](BasicBlock *OldBB) { 1182 // Clone the basic block and insert it before the new preheader. 1183 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent()); 1184 NewBB->moveBefore(LoopPH); 1185 1186 // Record this block and the mapping. 1187 NewBlocks.push_back(NewBB); 1188 VMap[OldBB] = NewBB; 1189 1190 return NewBB; 1191 }; 1192 1193 // We skip cloning blocks when they have a dominating succ that is not the 1194 // succ we are cloning for. 1195 auto SkipBlock = [&](BasicBlock *BB) { 1196 auto It = DominatingSucc.find(BB); 1197 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB; 1198 }; 1199 1200 // First, clone the preheader. 1201 auto *ClonedPH = CloneBlock(LoopPH); 1202 1203 // Then clone all the loop blocks, skipping the ones that aren't necessary. 1204 for (auto *LoopBB : L.blocks()) 1205 if (!SkipBlock(LoopBB)) 1206 CloneBlock(LoopBB); 1207 1208 // Split all the loop exit edges so that when we clone the exit blocks, if 1209 // any of the exit blocks are *also* a preheader for some other loop, we 1210 // don't create multiple predecessors entering the loop header. 1211 for (auto *ExitBB : ExitBlocks) { 1212 if (SkipBlock(ExitBB)) 1213 continue; 1214 1215 // When we are going to clone an exit, we don't need to clone all the 1216 // instructions in the exit block and we want to ensure we have an easy 1217 // place to merge the CFG, so split the exit first. This is always safe to 1218 // do because there cannot be any non-loop predecessors of a loop exit in 1219 // loop simplified form. 1220 auto *MergeBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU); 1221 1222 // Rearrange the names to make it easier to write test cases by having the 1223 // exit block carry the suffix rather than the merge block carrying the 1224 // suffix. 1225 MergeBB->takeName(ExitBB); 1226 ExitBB->setName(Twine(MergeBB->getName()) + ".split"); 1227 1228 // Now clone the original exit block. 1229 auto *ClonedExitBB = CloneBlock(ExitBB); 1230 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && 1231 "Exit block should have been split to have one successor!"); 1232 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && 1233 "Cloned exit block has the wrong successor!"); 1234 1235 // Remap any cloned instructions and create a merge phi node for them. 1236 for (auto ZippedInsts : llvm::zip_first( 1237 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())), 1238 llvm::make_range(ClonedExitBB->begin(), 1239 std::prev(ClonedExitBB->end())))) { 1240 Instruction &I = std::get<0>(ZippedInsts); 1241 Instruction &ClonedI = std::get<1>(ZippedInsts); 1242 1243 // The only instructions in the exit block should be PHI nodes and 1244 // potentially a landing pad. 1245 assert( 1246 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && 1247 "Bad instruction in exit block!"); 1248 // We should have a value map between the instruction and its clone. 1249 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!"); 1250 1251 // Forget SCEVs based on exit phis in case SCEV looked through the phi. 1252 if (SE && isa<PHINode>(I)) 1253 SE->forgetValue(&I); 1254 1255 BasicBlock::iterator InsertPt = MergeBB->getFirstInsertionPt(); 1256 1257 auto *MergePN = 1258 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi"); 1259 MergePN->insertBefore(InsertPt); 1260 MergePN->setDebugLoc(InsertPt->getDebugLoc()); 1261 I.replaceAllUsesWith(MergePN); 1262 MergePN->addIncoming(&I, ExitBB); 1263 MergePN->addIncoming(&ClonedI, ClonedExitBB); 1264 } 1265 } 1266 1267 // Rewrite the instructions in the cloned blocks to refer to the instructions 1268 // in the cloned blocks. We have to do this as a second pass so that we have 1269 // everything available. Also, we have inserted new instructions which may 1270 // include assume intrinsics, so we update the assumption cache while 1271 // processing this. 1272 Module *M = ClonedPH->getParent()->getParent(); 1273 for (auto *ClonedBB : NewBlocks) 1274 for (Instruction &I : *ClonedBB) { 1275 RemapDbgRecordRange(M, I.getDbgRecordRange(), VMap, 1276 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1277 RemapInstruction(&I, VMap, 1278 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1279 if (auto *II = dyn_cast<AssumeInst>(&I)) 1280 AC.registerAssumption(II); 1281 } 1282 1283 // Update any PHI nodes in the cloned successors of the skipped blocks to not 1284 // have spurious incoming values. 1285 for (auto *LoopBB : L.blocks()) 1286 if (SkipBlock(LoopBB)) 1287 for (auto *SuccBB : successors(LoopBB)) 1288 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB))) 1289 for (PHINode &PN : ClonedSuccBB->phis()) 1290 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false); 1291 1292 // Remove the cloned parent as a predecessor of any successor we ended up 1293 // cloning other than the unswitched one. 1294 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB)); 1295 for (auto *SuccBB : successors(ParentBB)) { 1296 if (SuccBB == UnswitchedSuccBB) 1297 continue; 1298 1299 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)); 1300 if (!ClonedSuccBB) 1301 continue; 1302 1303 ClonedSuccBB->removePredecessor(ClonedParentBB, 1304 /*KeepOneInputPHIs*/ true); 1305 } 1306 1307 // Replace the cloned branch with an unconditional branch to the cloned 1308 // unswitched successor. 1309 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB)); 1310 Instruction *ClonedTerminator = ClonedParentBB->getTerminator(); 1311 // Trivial Simplification. If Terminator is a conditional branch and 1312 // condition becomes dead - erase it. 1313 Value *ClonedConditionToErase = nullptr; 1314 if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator)) 1315 ClonedConditionToErase = BI->getCondition(); 1316 else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator)) 1317 ClonedConditionToErase = SI->getCondition(); 1318 1319 Instruction *BI = BranchInst::Create(ClonedSuccBB, ClonedParentBB); 1320 BI->setDebugLoc(ClonedTerminator->getDebugLoc()); 1321 ClonedTerminator->eraseFromParent(); 1322 1323 if (ClonedConditionToErase) 1324 RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr, 1325 MSSAU); 1326 1327 // If there are duplicate entries in the PHI nodes because of multiple edges 1328 // to the unswitched successor, we need to nuke all but one as we replaced it 1329 // with a direct branch. 1330 for (PHINode &PN : ClonedSuccBB->phis()) { 1331 bool Found = false; 1332 // Loop over the incoming operands backwards so we can easily delete as we 1333 // go without invalidating the index. 1334 for (int i = PN.getNumOperands() - 1; i >= 0; --i) { 1335 if (PN.getIncomingBlock(i) != ClonedParentBB) 1336 continue; 1337 if (!Found) { 1338 Found = true; 1339 continue; 1340 } 1341 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false); 1342 } 1343 } 1344 1345 // Record the domtree updates for the new blocks. 1346 SmallPtrSet<BasicBlock *, 4> SuccSet; 1347 for (auto *ClonedBB : NewBlocks) { 1348 for (auto *SuccBB : successors(ClonedBB)) 1349 if (SuccSet.insert(SuccBB).second) 1350 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB}); 1351 SuccSet.clear(); 1352 } 1353 1354 return ClonedPH; 1355 } 1356 1357 /// Recursively clone the specified loop and all of its children. 1358 /// 1359 /// The target parent loop for the clone should be provided, or can be null if 1360 /// the clone is a top-level loop. While cloning, all the blocks are mapped 1361 /// with the provided value map. The entire original loop must be present in 1362 /// the value map. The cloned loop is returned. 1363 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, 1364 const ValueToValueMapTy &VMap, LoopInfo &LI) { 1365 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) { 1366 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!"); 1367 ClonedL.reserveBlocks(OrigL.getNumBlocks()); 1368 for (auto *BB : OrigL.blocks()) { 1369 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB)); 1370 ClonedL.addBlockEntry(ClonedBB); 1371 if (LI.getLoopFor(BB) == &OrigL) 1372 LI.changeLoopFor(ClonedBB, &ClonedL); 1373 } 1374 }; 1375 1376 // We specially handle the first loop because it may get cloned into 1377 // a different parent and because we most commonly are cloning leaf loops. 1378 Loop *ClonedRootL = LI.AllocateLoop(); 1379 if (RootParentL) 1380 RootParentL->addChildLoop(ClonedRootL); 1381 else 1382 LI.addTopLevelLoop(ClonedRootL); 1383 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL); 1384 1385 if (OrigRootL.isInnermost()) 1386 return ClonedRootL; 1387 1388 // If we have a nest, we can quickly clone the entire loop nest using an 1389 // iterative approach because it is a tree. We keep the cloned parent in the 1390 // data structure to avoid repeatedly querying through a map to find it. 1391 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone; 1392 // Build up the loops to clone in reverse order as we'll clone them from the 1393 // back. 1394 for (Loop *ChildL : llvm::reverse(OrigRootL)) 1395 LoopsToClone.push_back({ClonedRootL, ChildL}); 1396 do { 1397 Loop *ClonedParentL, *L; 1398 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val(); 1399 Loop *ClonedL = LI.AllocateLoop(); 1400 ClonedParentL->addChildLoop(ClonedL); 1401 AddClonedBlocksToLoop(*L, *ClonedL); 1402 for (Loop *ChildL : llvm::reverse(*L)) 1403 LoopsToClone.push_back({ClonedL, ChildL}); 1404 } while (!LoopsToClone.empty()); 1405 1406 return ClonedRootL; 1407 } 1408 1409 /// Build the cloned loops of an original loop from unswitching. 1410 /// 1411 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial 1412 /// operation. We need to re-verify that there even is a loop (as the backedge 1413 /// may not have been cloned), and even if there are remaining backedges the 1414 /// backedge set may be different. However, we know that each child loop is 1415 /// undisturbed, we only need to find where to place each child loop within 1416 /// either any parent loop or within a cloned version of the original loop. 1417 /// 1418 /// Because child loops may end up cloned outside of any cloned version of the 1419 /// original loop, multiple cloned sibling loops may be created. All of them 1420 /// are returned so that the newly introduced loop nest roots can be 1421 /// identified. 1422 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks, 1423 const ValueToValueMapTy &VMap, LoopInfo &LI, 1424 SmallVectorImpl<Loop *> &NonChildClonedLoops) { 1425 Loop *ClonedL = nullptr; 1426 1427 auto *OrigPH = OrigL.getLoopPreheader(); 1428 auto *OrigHeader = OrigL.getHeader(); 1429 1430 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH)); 1431 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader)); 1432 1433 // We need to know the loops of the cloned exit blocks to even compute the 1434 // accurate parent loop. If we only clone exits to some parent of the 1435 // original parent, we want to clone into that outer loop. We also keep track 1436 // of the loops that our cloned exit blocks participate in. 1437 Loop *ParentL = nullptr; 1438 SmallVector<BasicBlock *, 4> ClonedExitsInLoops; 1439 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap; 1440 ClonedExitsInLoops.reserve(ExitBlocks.size()); 1441 for (auto *ExitBB : ExitBlocks) 1442 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB))) 1443 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1444 ExitLoopMap[ClonedExitBB] = ExitL; 1445 ClonedExitsInLoops.push_back(ClonedExitBB); 1446 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1447 ParentL = ExitL; 1448 } 1449 assert((!ParentL || ParentL == OrigL.getParentLoop() || 1450 ParentL->contains(OrigL.getParentLoop())) && 1451 "The computed parent loop should always contain (or be) the parent of " 1452 "the original loop."); 1453 1454 // We build the set of blocks dominated by the cloned header from the set of 1455 // cloned blocks out of the original loop. While not all of these will 1456 // necessarily be in the cloned loop, it is enough to establish that they 1457 // aren't in unreachable cycles, etc. 1458 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks; 1459 for (auto *BB : OrigL.blocks()) 1460 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB))) 1461 ClonedLoopBlocks.insert(ClonedBB); 1462 1463 // Rebuild the set of blocks that will end up in the cloned loop. We may have 1464 // skipped cloning some region of this loop which can in turn skip some of 1465 // the backedges so we have to rebuild the blocks in the loop based on the 1466 // backedges that remain after cloning. 1467 SmallVector<BasicBlock *, 16> Worklist; 1468 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop; 1469 for (auto *Pred : predecessors(ClonedHeader)) { 1470 // The only possible non-loop header predecessor is the preheader because 1471 // we know we cloned the loop in simplified form. 1472 if (Pred == ClonedPH) 1473 continue; 1474 1475 // Because the loop was in simplified form, the only non-loop predecessor 1476 // should be the preheader. 1477 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop " 1478 "header other than the preheader " 1479 "that is not part of the loop!"); 1480 1481 // Insert this block into the loop set and on the first visit (and if it 1482 // isn't the header we're currently walking) put it into the worklist to 1483 // recurse through. 1484 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader) 1485 Worklist.push_back(Pred); 1486 } 1487 1488 // If we had any backedges then there *is* a cloned loop. Put the header into 1489 // the loop set and then walk the worklist backwards to find all the blocks 1490 // that remain within the loop after cloning. 1491 if (!BlocksInClonedLoop.empty()) { 1492 BlocksInClonedLoop.insert(ClonedHeader); 1493 1494 while (!Worklist.empty()) { 1495 BasicBlock *BB = Worklist.pop_back_val(); 1496 assert(BlocksInClonedLoop.count(BB) && 1497 "Didn't put block into the loop set!"); 1498 1499 // Insert any predecessors that are in the possible set into the cloned 1500 // set, and if the insert is successful, add them to the worklist. Note 1501 // that we filter on the blocks that are definitely reachable via the 1502 // backedge to the loop header so we may prune out dead code within the 1503 // cloned loop. 1504 for (auto *Pred : predecessors(BB)) 1505 if (ClonedLoopBlocks.count(Pred) && 1506 BlocksInClonedLoop.insert(Pred).second) 1507 Worklist.push_back(Pred); 1508 } 1509 1510 ClonedL = LI.AllocateLoop(); 1511 if (ParentL) { 1512 ParentL->addBasicBlockToLoop(ClonedPH, LI); 1513 ParentL->addChildLoop(ClonedL); 1514 } else { 1515 LI.addTopLevelLoop(ClonedL); 1516 } 1517 NonChildClonedLoops.push_back(ClonedL); 1518 1519 ClonedL->reserveBlocks(BlocksInClonedLoop.size()); 1520 // We don't want to just add the cloned loop blocks based on how we 1521 // discovered them. The original order of blocks was carefully built in 1522 // a way that doesn't rely on predecessor ordering. Rather than re-invent 1523 // that logic, we just re-walk the original blocks (and those of the child 1524 // loops) and filter them as we add them into the cloned loop. 1525 for (auto *BB : OrigL.blocks()) { 1526 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)); 1527 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB)) 1528 continue; 1529 1530 // Directly add the blocks that are only in this loop. 1531 if (LI.getLoopFor(BB) == &OrigL) { 1532 ClonedL->addBasicBlockToLoop(ClonedBB, LI); 1533 continue; 1534 } 1535 1536 // We want to manually add it to this loop and parents. 1537 // Registering it with LoopInfo will happen when we clone the top 1538 // loop for this block. 1539 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop()) 1540 PL->addBlockEntry(ClonedBB); 1541 } 1542 1543 // Now add each child loop whose header remains within the cloned loop. All 1544 // of the blocks within the loop must satisfy the same constraints as the 1545 // header so once we pass the header checks we can just clone the entire 1546 // child loop nest. 1547 for (Loop *ChildL : OrigL) { 1548 auto *ClonedChildHeader = 1549 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1550 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader)) 1551 continue; 1552 1553 #ifndef NDEBUG 1554 // We should never have a cloned child loop header but fail to have 1555 // all of the blocks for that child loop. 1556 for (auto *ChildLoopBB : ChildL->blocks()) 1557 assert(BlocksInClonedLoop.count( 1558 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && 1559 "Child cloned loop has a header within the cloned outer " 1560 "loop but not all of its blocks!"); 1561 #endif 1562 1563 cloneLoopNest(*ChildL, ClonedL, VMap, LI); 1564 } 1565 } 1566 1567 // Now that we've handled all the components of the original loop that were 1568 // cloned into a new loop, we still need to handle anything from the original 1569 // loop that wasn't in a cloned loop. 1570 1571 // Figure out what blocks are left to place within any loop nest containing 1572 // the unswitched loop. If we never formed a loop, the cloned PH is one of 1573 // them. 1574 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet; 1575 if (BlocksInClonedLoop.empty()) 1576 UnloopedBlockSet.insert(ClonedPH); 1577 for (auto *ClonedBB : ClonedLoopBlocks) 1578 if (!BlocksInClonedLoop.count(ClonedBB)) 1579 UnloopedBlockSet.insert(ClonedBB); 1580 1581 // Copy the cloned exits and sort them in ascending loop depth, we'll work 1582 // backwards across these to process them inside out. The order shouldn't 1583 // matter as we're just trying to build up the map from inside-out; we use 1584 // the map in a more stably ordered way below. 1585 auto OrderedClonedExitsInLoops = ClonedExitsInLoops; 1586 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { 1587 return ExitLoopMap.lookup(LHS)->getLoopDepth() < 1588 ExitLoopMap.lookup(RHS)->getLoopDepth(); 1589 }); 1590 1591 // Populate the existing ExitLoopMap with everything reachable from each 1592 // exit, starting from the inner most exit. 1593 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) { 1594 assert(Worklist.empty() && "Didn't clear worklist!"); 1595 1596 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val(); 1597 Loop *ExitL = ExitLoopMap.lookup(ExitBB); 1598 1599 // Walk the CFG back until we hit the cloned PH adding everything reachable 1600 // and in the unlooped set to this exit block's loop. 1601 Worklist.push_back(ExitBB); 1602 do { 1603 BasicBlock *BB = Worklist.pop_back_val(); 1604 // We can stop recursing at the cloned preheader (if we get there). 1605 if (BB == ClonedPH) 1606 continue; 1607 1608 for (BasicBlock *PredBB : predecessors(BB)) { 1609 // If this pred has already been moved to our set or is part of some 1610 // (inner) loop, no update needed. 1611 if (!UnloopedBlockSet.erase(PredBB)) { 1612 assert( 1613 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && 1614 "Predecessor not mapped to a loop!"); 1615 continue; 1616 } 1617 1618 // We just insert into the loop set here. We'll add these blocks to the 1619 // exit loop after we build up the set in an order that doesn't rely on 1620 // predecessor order (which in turn relies on use list order). 1621 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second; 1622 (void)Inserted; 1623 assert(Inserted && "Should only visit an unlooped block once!"); 1624 1625 // And recurse through to its predecessors. 1626 Worklist.push_back(PredBB); 1627 } 1628 } while (!Worklist.empty()); 1629 } 1630 1631 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned 1632 // blocks to their outer loops, walk the cloned blocks and the cloned exits 1633 // in their original order adding them to the correct loop. 1634 1635 // We need a stable insertion order. We use the order of the original loop 1636 // order and map into the correct parent loop. 1637 for (auto *BB : llvm::concat<BasicBlock *const>( 1638 ArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops)) 1639 if (Loop *OuterL = ExitLoopMap.lookup(BB)) 1640 OuterL->addBasicBlockToLoop(BB, LI); 1641 1642 #ifndef NDEBUG 1643 for (auto &BBAndL : ExitLoopMap) { 1644 auto *BB = BBAndL.first; 1645 auto *OuterL = BBAndL.second; 1646 assert(LI.getLoopFor(BB) == OuterL && 1647 "Failed to put all blocks into outer loops!"); 1648 } 1649 #endif 1650 1651 // Now that all the blocks are placed into the correct containing loop in the 1652 // absence of child loops, find all the potentially cloned child loops and 1653 // clone them into whatever outer loop we placed their header into. 1654 for (Loop *ChildL : OrigL) { 1655 auto *ClonedChildHeader = 1656 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1657 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader)) 1658 continue; 1659 1660 #ifndef NDEBUG 1661 for (auto *ChildLoopBB : ChildL->blocks()) 1662 assert(VMap.count(ChildLoopBB) && 1663 "Cloned a child loop header but not all of that loops blocks!"); 1664 #endif 1665 1666 NonChildClonedLoops.push_back(cloneLoopNest( 1667 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI)); 1668 } 1669 } 1670 1671 static void 1672 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1673 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, 1674 DominatorTree &DT, MemorySSAUpdater *MSSAU) { 1675 // Find all the dead clones, and remove them from their successors. 1676 SmallVector<BasicBlock *, 16> DeadBlocks; 1677 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks)) 1678 for (const auto &VMap : VMaps) 1679 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB))) 1680 if (!DT.isReachableFromEntry(ClonedBB)) { 1681 for (BasicBlock *SuccBB : successors(ClonedBB)) 1682 SuccBB->removePredecessor(ClonedBB); 1683 DeadBlocks.push_back(ClonedBB); 1684 } 1685 1686 // Remove all MemorySSA in the dead blocks 1687 if (MSSAU) { 1688 SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(), 1689 DeadBlocks.end()); 1690 MSSAU->removeBlocks(DeadBlockSet); 1691 } 1692 1693 // Drop any remaining references to break cycles. 1694 for (BasicBlock *BB : DeadBlocks) 1695 BB->dropAllReferences(); 1696 // Erase them from the IR. 1697 for (BasicBlock *BB : DeadBlocks) 1698 BB->eraseFromParent(); 1699 } 1700 1701 static void deleteDeadBlocksFromLoop(Loop &L, 1702 SmallVectorImpl<BasicBlock *> &ExitBlocks, 1703 DominatorTree &DT, LoopInfo &LI, 1704 MemorySSAUpdater *MSSAU, 1705 ScalarEvolution *SE, 1706 LPMUpdater &LoopUpdater) { 1707 // Find all the dead blocks tied to this loop, and remove them from their 1708 // successors. 1709 SmallSetVector<BasicBlock *, 8> DeadBlockSet; 1710 1711 // Start with loop/exit blocks and get a transitive closure of reachable dead 1712 // blocks. 1713 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(), 1714 ExitBlocks.end()); 1715 DeathCandidates.append(L.blocks().begin(), L.blocks().end()); 1716 while (!DeathCandidates.empty()) { 1717 auto *BB = DeathCandidates.pop_back_val(); 1718 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) { 1719 for (BasicBlock *SuccBB : successors(BB)) { 1720 SuccBB->removePredecessor(BB); 1721 DeathCandidates.push_back(SuccBB); 1722 } 1723 DeadBlockSet.insert(BB); 1724 } 1725 } 1726 1727 // Remove all MemorySSA in the dead blocks 1728 if (MSSAU) 1729 MSSAU->removeBlocks(DeadBlockSet); 1730 1731 // Filter out the dead blocks from the exit blocks list so that it can be 1732 // used in the caller. 1733 llvm::erase_if(ExitBlocks, 1734 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1735 1736 // Walk from this loop up through its parents removing all of the dead blocks. 1737 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) { 1738 for (auto *BB : DeadBlockSet) 1739 ParentL->getBlocksSet().erase(BB); 1740 llvm::erase_if(ParentL->getBlocksVector(), 1741 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1742 } 1743 1744 // Now delete the dead child loops. This raw delete will clear them 1745 // recursively. 1746 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) { 1747 if (!DeadBlockSet.count(ChildL->getHeader())) 1748 return false; 1749 1750 assert(llvm::all_of(ChildL->blocks(), 1751 [&](BasicBlock *ChildBB) { 1752 return DeadBlockSet.count(ChildBB); 1753 }) && 1754 "If the child loop header is dead all blocks in the child loop must " 1755 "be dead as well!"); 1756 LoopUpdater.markLoopAsDeleted(*ChildL, ChildL->getName()); 1757 if (SE) 1758 SE->forgetBlockAndLoopDispositions(); 1759 LI.destroy(ChildL); 1760 return true; 1761 }); 1762 1763 // Remove the loop mappings for the dead blocks and drop all the references 1764 // from these blocks to others to handle cyclic references as we start 1765 // deleting the blocks themselves. 1766 for (auto *BB : DeadBlockSet) { 1767 // Check that the dominator tree has already been updated. 1768 assert(!DT.getNode(BB) && "Should already have cleared domtree!"); 1769 LI.changeLoopFor(BB, nullptr); 1770 // Drop all uses of the instructions to make sure we won't have dangling 1771 // uses in other blocks. 1772 for (auto &I : *BB) 1773 if (!I.use_empty()) 1774 I.replaceAllUsesWith(PoisonValue::get(I.getType())); 1775 BB->dropAllReferences(); 1776 } 1777 1778 // Actually delete the blocks now that they've been fully unhooked from the 1779 // IR. 1780 for (auto *BB : DeadBlockSet) 1781 BB->eraseFromParent(); 1782 } 1783 1784 /// Recompute the set of blocks in a loop after unswitching. 1785 /// 1786 /// This walks from the original headers predecessors to rebuild the loop. We 1787 /// take advantage of the fact that new blocks can't have been added, and so we 1788 /// filter by the original loop's blocks. This also handles potentially 1789 /// unreachable code that we don't want to explore but might be found examining 1790 /// the predecessors of the header. 1791 /// 1792 /// If the original loop is no longer a loop, this will return an empty set. If 1793 /// it remains a loop, all the blocks within it will be added to the set 1794 /// (including those blocks in inner loops). 1795 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L, 1796 LoopInfo &LI) { 1797 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet; 1798 1799 auto *PH = L.getLoopPreheader(); 1800 auto *Header = L.getHeader(); 1801 1802 // A worklist to use while walking backwards from the header. 1803 SmallVector<BasicBlock *, 16> Worklist; 1804 1805 // First walk the predecessors of the header to find the backedges. This will 1806 // form the basis of our walk. 1807 for (auto *Pred : predecessors(Header)) { 1808 // Skip the preheader. 1809 if (Pred == PH) 1810 continue; 1811 1812 // Because the loop was in simplified form, the only non-loop predecessor 1813 // is the preheader. 1814 assert(L.contains(Pred) && "Found a predecessor of the loop header other " 1815 "than the preheader that is not part of the " 1816 "loop!"); 1817 1818 // Insert this block into the loop set and on the first visit and, if it 1819 // isn't the header we're currently walking, put it into the worklist to 1820 // recurse through. 1821 if (LoopBlockSet.insert(Pred).second && Pred != Header) 1822 Worklist.push_back(Pred); 1823 } 1824 1825 // If no backedges were found, we're done. 1826 if (LoopBlockSet.empty()) 1827 return LoopBlockSet; 1828 1829 // We found backedges, recurse through them to identify the loop blocks. 1830 while (!Worklist.empty()) { 1831 BasicBlock *BB = Worklist.pop_back_val(); 1832 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!"); 1833 1834 // No need to walk past the header. 1835 if (BB == Header) 1836 continue; 1837 1838 // Because we know the inner loop structure remains valid we can use the 1839 // loop structure to jump immediately across the entire nested loop. 1840 // Further, because it is in loop simplified form, we can directly jump 1841 // to its preheader afterward. 1842 if (Loop *InnerL = LI.getLoopFor(BB)) 1843 if (InnerL != &L) { 1844 assert(L.contains(InnerL) && 1845 "Should not reach a loop *outside* this loop!"); 1846 // The preheader is the only possible predecessor of the loop so 1847 // insert it into the set and check whether it was already handled. 1848 auto *InnerPH = InnerL->getLoopPreheader(); 1849 assert(L.contains(InnerPH) && "Cannot contain an inner loop block " 1850 "but not contain the inner loop " 1851 "preheader!"); 1852 if (!LoopBlockSet.insert(InnerPH).second) 1853 // The only way to reach the preheader is through the loop body 1854 // itself so if it has been visited the loop is already handled. 1855 continue; 1856 1857 // Insert all of the blocks (other than those already present) into 1858 // the loop set. We expect at least the block that led us to find the 1859 // inner loop to be in the block set, but we may also have other loop 1860 // blocks if they were already enqueued as predecessors of some other 1861 // outer loop block. 1862 for (auto *InnerBB : InnerL->blocks()) { 1863 if (InnerBB == BB) { 1864 assert(LoopBlockSet.count(InnerBB) && 1865 "Block should already be in the set!"); 1866 continue; 1867 } 1868 1869 LoopBlockSet.insert(InnerBB); 1870 } 1871 1872 // Add the preheader to the worklist so we will continue past the 1873 // loop body. 1874 Worklist.push_back(InnerPH); 1875 continue; 1876 } 1877 1878 // Insert any predecessors that were in the original loop into the new 1879 // set, and if the insert is successful, add them to the worklist. 1880 for (auto *Pred : predecessors(BB)) 1881 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second) 1882 Worklist.push_back(Pred); 1883 } 1884 1885 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!"); 1886 1887 // We've found all the blocks participating in the loop, return our completed 1888 // set. 1889 return LoopBlockSet; 1890 } 1891 1892 /// Rebuild a loop after unswitching removes some subset of blocks and edges. 1893 /// 1894 /// The removal may have removed some child loops entirely but cannot have 1895 /// disturbed any remaining child loops. However, they may need to be hoisted 1896 /// to the parent loop (or to be top-level loops). The original loop may be 1897 /// completely removed. 1898 /// 1899 /// The sibling loops resulting from this update are returned. If the original 1900 /// loop remains a valid loop, it will be the first entry in this list with all 1901 /// of the newly sibling loops following it. 1902 /// 1903 /// Returns true if the loop remains a loop after unswitching, and false if it 1904 /// is no longer a loop after unswitching (and should not continue to be 1905 /// referenced). 1906 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1907 LoopInfo &LI, 1908 SmallVectorImpl<Loop *> &HoistedLoops, 1909 ScalarEvolution *SE) { 1910 auto *PH = L.getLoopPreheader(); 1911 1912 // Compute the actual parent loop from the exit blocks. Because we may have 1913 // pruned some exits the loop may be different from the original parent. 1914 Loop *ParentL = nullptr; 1915 SmallVector<Loop *, 4> ExitLoops; 1916 SmallVector<BasicBlock *, 4> ExitsInLoops; 1917 ExitsInLoops.reserve(ExitBlocks.size()); 1918 for (auto *ExitBB : ExitBlocks) 1919 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1920 ExitLoops.push_back(ExitL); 1921 ExitsInLoops.push_back(ExitBB); 1922 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1923 ParentL = ExitL; 1924 } 1925 1926 // Recompute the blocks participating in this loop. This may be empty if it 1927 // is no longer a loop. 1928 auto LoopBlockSet = recomputeLoopBlockSet(L, LI); 1929 1930 // If we still have a loop, we need to re-set the loop's parent as the exit 1931 // block set changing may have moved it within the loop nest. Note that this 1932 // can only happen when this loop has a parent as it can only hoist the loop 1933 // *up* the nest. 1934 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) { 1935 // Remove this loop's (original) blocks from all of the intervening loops. 1936 for (Loop *IL = L.getParentLoop(); IL != ParentL; 1937 IL = IL->getParentLoop()) { 1938 IL->getBlocksSet().erase(PH); 1939 for (auto *BB : L.blocks()) 1940 IL->getBlocksSet().erase(BB); 1941 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) { 1942 return BB == PH || L.contains(BB); 1943 }); 1944 } 1945 1946 LI.changeLoopFor(PH, ParentL); 1947 L.getParentLoop()->removeChildLoop(&L); 1948 if (ParentL) 1949 ParentL->addChildLoop(&L); 1950 else 1951 LI.addTopLevelLoop(&L); 1952 } 1953 1954 // Now we update all the blocks which are no longer within the loop. 1955 auto &Blocks = L.getBlocksVector(); 1956 auto BlocksSplitI = 1957 LoopBlockSet.empty() 1958 ? Blocks.begin() 1959 : std::stable_partition( 1960 Blocks.begin(), Blocks.end(), 1961 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); }); 1962 1963 // Before we erase the list of unlooped blocks, build a set of them. 1964 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end()); 1965 if (LoopBlockSet.empty()) 1966 UnloopedBlocks.insert(PH); 1967 1968 // Now erase these blocks from the loop. 1969 for (auto *BB : make_range(BlocksSplitI, Blocks.end())) 1970 L.getBlocksSet().erase(BB); 1971 Blocks.erase(BlocksSplitI, Blocks.end()); 1972 1973 // Sort the exits in ascending loop depth, we'll work backwards across these 1974 // to process them inside out. 1975 llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { 1976 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS); 1977 }); 1978 1979 // We'll build up a set for each exit loop. 1980 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks; 1981 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop. 1982 1983 auto RemoveUnloopedBlocksFromLoop = 1984 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) { 1985 for (auto *BB : UnloopedBlocks) 1986 L.getBlocksSet().erase(BB); 1987 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) { 1988 return UnloopedBlocks.count(BB); 1989 }); 1990 }; 1991 1992 SmallVector<BasicBlock *, 16> Worklist; 1993 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) { 1994 assert(Worklist.empty() && "Didn't clear worklist!"); 1995 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!"); 1996 1997 // Grab the next exit block, in decreasing loop depth order. 1998 BasicBlock *ExitBB = ExitsInLoops.pop_back_val(); 1999 Loop &ExitL = *LI.getLoopFor(ExitBB); 2000 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!"); 2001 2002 // Erase all of the unlooped blocks from the loops between the previous 2003 // exit loop and this exit loop. This works because the ExitInLoops list is 2004 // sorted in increasing order of loop depth and thus we visit loops in 2005 // decreasing order of loop depth. 2006 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop()) 2007 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 2008 2009 // Walk the CFG back until we hit the cloned PH adding everything reachable 2010 // and in the unlooped set to this exit block's loop. 2011 Worklist.push_back(ExitBB); 2012 do { 2013 BasicBlock *BB = Worklist.pop_back_val(); 2014 // We can stop recursing at the cloned preheader (if we get there). 2015 if (BB == PH) 2016 continue; 2017 2018 for (BasicBlock *PredBB : predecessors(BB)) { 2019 // If this pred has already been moved to our set or is part of some 2020 // (inner) loop, no update needed. 2021 if (!UnloopedBlocks.erase(PredBB)) { 2022 assert((NewExitLoopBlocks.count(PredBB) || 2023 ExitL.contains(LI.getLoopFor(PredBB))) && 2024 "Predecessor not in a nested loop (or already visited)!"); 2025 continue; 2026 } 2027 2028 // We just insert into the loop set here. We'll add these blocks to the 2029 // exit loop after we build up the set in a deterministic order rather 2030 // than the predecessor-influenced visit order. 2031 bool Inserted = NewExitLoopBlocks.insert(PredBB).second; 2032 (void)Inserted; 2033 assert(Inserted && "Should only visit an unlooped block once!"); 2034 2035 // And recurse through to its predecessors. 2036 Worklist.push_back(PredBB); 2037 } 2038 } while (!Worklist.empty()); 2039 2040 // If blocks in this exit loop were directly part of the original loop (as 2041 // opposed to a child loop) update the map to point to this exit loop. This 2042 // just updates a map and so the fact that the order is unstable is fine. 2043 for (auto *BB : NewExitLoopBlocks) 2044 if (Loop *BBL = LI.getLoopFor(BB)) 2045 if (BBL == &L || !L.contains(BBL)) 2046 LI.changeLoopFor(BB, &ExitL); 2047 2048 // We will remove the remaining unlooped blocks from this loop in the next 2049 // iteration or below. 2050 NewExitLoopBlocks.clear(); 2051 } 2052 2053 // Any remaining unlooped blocks are no longer part of any loop unless they 2054 // are part of some child loop. 2055 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop()) 2056 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 2057 for (auto *BB : UnloopedBlocks) 2058 if (Loop *BBL = LI.getLoopFor(BB)) 2059 if (BBL == &L || !L.contains(BBL)) 2060 LI.changeLoopFor(BB, nullptr); 2061 2062 // Sink all the child loops whose headers are no longer in the loop set to 2063 // the parent (or to be top level loops). We reach into the loop and directly 2064 // update its subloop vector to make this batch update efficient. 2065 auto &SubLoops = L.getSubLoopsVector(); 2066 auto SubLoopsSplitI = 2067 LoopBlockSet.empty() 2068 ? SubLoops.begin() 2069 : std::stable_partition( 2070 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) { 2071 return LoopBlockSet.count(SubL->getHeader()); 2072 }); 2073 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) { 2074 HoistedLoops.push_back(HoistedL); 2075 HoistedL->setParentLoop(nullptr); 2076 2077 // To compute the new parent of this hoisted loop we look at where we 2078 // placed the preheader above. We can't lookup the header itself because we 2079 // retained the mapping from the header to the hoisted loop. But the 2080 // preheader and header should have the exact same new parent computed 2081 // based on the set of exit blocks from the original loop as the preheader 2082 // is a predecessor of the header and so reached in the reverse walk. And 2083 // because the loops were all in simplified form the preheader of the 2084 // hoisted loop can't be part of some *other* loop. 2085 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader())) 2086 NewParentL->addChildLoop(HoistedL); 2087 else 2088 LI.addTopLevelLoop(HoistedL); 2089 } 2090 SubLoops.erase(SubLoopsSplitI, SubLoops.end()); 2091 2092 // Actually delete the loop if nothing remained within it. 2093 if (Blocks.empty()) { 2094 assert(SubLoops.empty() && 2095 "Failed to remove all subloops from the original loop!"); 2096 if (Loop *ParentL = L.getParentLoop()) 2097 ParentL->removeChildLoop(llvm::find(*ParentL, &L)); 2098 else 2099 LI.removeLoop(llvm::find(LI, &L)); 2100 // markLoopAsDeleted for L should be triggered by the caller (it is 2101 // typically done within postUnswitch). 2102 if (SE) 2103 SE->forgetBlockAndLoopDispositions(); 2104 LI.destroy(&L); 2105 return false; 2106 } 2107 2108 return true; 2109 } 2110 2111 /// Helper to visit a dominator subtree, invoking a callable on each node. 2112 /// 2113 /// Returning false at any point will stop walking past that node of the tree. 2114 template <typename CallableT> 2115 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) { 2116 SmallVector<DomTreeNode *, 4> DomWorklist; 2117 DomWorklist.push_back(DT[BB]); 2118 #ifndef NDEBUG 2119 SmallPtrSet<DomTreeNode *, 4> Visited; 2120 Visited.insert(DT[BB]); 2121 #endif 2122 do { 2123 DomTreeNode *N = DomWorklist.pop_back_val(); 2124 2125 // Visit this node. 2126 if (!Callable(N->getBlock())) 2127 continue; 2128 2129 // Accumulate the child nodes. 2130 for (DomTreeNode *ChildN : *N) { 2131 assert(Visited.insert(ChildN).second && 2132 "Cannot visit a node twice when walking a tree!"); 2133 DomWorklist.push_back(ChildN); 2134 } 2135 } while (!DomWorklist.empty()); 2136 } 2137 2138 void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName, 2139 bool CurrentLoopValid, bool PartiallyInvariant, 2140 bool InjectedCondition, ArrayRef<Loop *> NewLoops) { 2141 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2142 if (!NewLoops.empty()) 2143 U.addSiblingLoops(NewLoops); 2144 2145 // If the current loop remains valid, we should revisit it to catch any 2146 // other unswitch opportunities. Otherwise, we need to mark it as deleted. 2147 if (CurrentLoopValid) { 2148 if (PartiallyInvariant) { 2149 // Mark the new loop as partially unswitched, to avoid unswitching on 2150 // the same condition again. 2151 auto &Context = L.getHeader()->getContext(); 2152 MDNode *DisableUnswitchMD = MDNode::get( 2153 Context, 2154 MDString::get(Context, "llvm.loop.unswitch.partial.disable")); 2155 MDNode *NewLoopID = makePostTransformationMetadata( 2156 Context, L.getLoopID(), {"llvm.loop.unswitch.partial"}, 2157 {DisableUnswitchMD}); 2158 L.setLoopID(NewLoopID); 2159 } else if (InjectedCondition) { 2160 // Do the same for injection of invariant conditions. 2161 auto &Context = L.getHeader()->getContext(); 2162 MDNode *DisableUnswitchMD = MDNode::get( 2163 Context, 2164 MDString::get(Context, "llvm.loop.unswitch.injection.disable")); 2165 MDNode *NewLoopID = makePostTransformationMetadata( 2166 Context, L.getLoopID(), {"llvm.loop.unswitch.injection"}, 2167 {DisableUnswitchMD}); 2168 L.setLoopID(NewLoopID); 2169 } else 2170 U.revisitCurrentLoop(); 2171 } else 2172 U.markLoopAsDeleted(L, LoopName); 2173 } 2174 2175 static void unswitchNontrivialInvariants( 2176 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants, 2177 IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI, 2178 AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU, 2179 LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) { 2180 auto *ParentBB = TI.getParent(); 2181 BranchInst *BI = dyn_cast<BranchInst>(&TI); 2182 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI); 2183 2184 // Save the current loop name in a variable so that we can report it even 2185 // after it has been deleted. 2186 std::string LoopName(L.getName()); 2187 2188 // We can only unswitch switches, conditional branches with an invariant 2189 // condition, or combining invariant conditions with an instruction or 2190 // partially invariant instructions. 2191 assert((SI || (BI && BI->isConditional())) && 2192 "Can only unswitch switches and conditional branch!"); 2193 bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty(); 2194 bool FullUnswitch = 2195 SI || (skipTrivialSelect(BI->getCondition()) == Invariants[0] && 2196 !PartiallyInvariant); 2197 if (FullUnswitch) 2198 assert(Invariants.size() == 1 && 2199 "Cannot have other invariants with full unswitching!"); 2200 else 2201 assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) && 2202 "Partial unswitching requires an instruction as the condition!"); 2203 2204 if (MSSAU && VerifyMemorySSA) 2205 MSSAU->getMemorySSA()->verifyMemorySSA(); 2206 2207 // Constant and BBs tracking the cloned and continuing successor. When we are 2208 // unswitching the entire condition, this can just be trivially chosen to 2209 // unswitch towards `true`. However, when we are unswitching a set of 2210 // invariants combined with `and` or `or` or partially invariant instructions, 2211 // the combining operation determines the best direction to unswitch: we want 2212 // to unswitch the direction that will collapse the branch. 2213 bool Direction = true; 2214 int ClonedSucc = 0; 2215 if (!FullUnswitch) { 2216 Value *Cond = skipTrivialSelect(BI->getCondition()); 2217 (void)Cond; 2218 assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) || 2219 PartiallyInvariant) && 2220 "Only `or`, `and`, an `select`, partially invariant instructions " 2221 "can combine invariants being unswitched."); 2222 if (!match(Cond, m_LogicalOr())) { 2223 if (match(Cond, m_LogicalAnd()) || 2224 (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) { 2225 Direction = false; 2226 ClonedSucc = 1; 2227 } 2228 } 2229 } 2230 2231 BasicBlock *RetainedSuccBB = 2232 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest(); 2233 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs; 2234 if (BI) 2235 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc)); 2236 else 2237 for (auto Case : SI->cases()) 2238 if (Case.getCaseSuccessor() != RetainedSuccBB) 2239 UnswitchedSuccBBs.insert(Case.getCaseSuccessor()); 2240 2241 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) && 2242 "Should not unswitch the same successor we are retaining!"); 2243 2244 // The branch should be in this exact loop. Any inner loop's invariant branch 2245 // should be handled by unswitching that inner loop. The caller of this 2246 // routine should filter out any candidates that remain (but were skipped for 2247 // whatever reason). 2248 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!"); 2249 2250 // Compute the parent loop now before we start hacking on things. 2251 Loop *ParentL = L.getParentLoop(); 2252 // Get blocks in RPO order for MSSA update, before changing the CFG. 2253 LoopBlocksRPO LBRPO(&L); 2254 if (MSSAU) 2255 LBRPO.perform(&LI); 2256 2257 // Compute the outer-most loop containing one of our exit blocks. This is the 2258 // furthest up our loopnest which can be mutated, which we will use below to 2259 // update things. 2260 Loop *OuterExitL = &L; 2261 SmallVector<BasicBlock *, 4> ExitBlocks; 2262 L.getUniqueExitBlocks(ExitBlocks); 2263 for (auto *ExitBB : ExitBlocks) { 2264 // ExitBB can be an exit block for several levels in the loop nest. Make 2265 // sure we find the top most. 2266 Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI); 2267 if (!NewOuterExitL) { 2268 // We exited the entire nest with this block, so we're done. 2269 OuterExitL = nullptr; 2270 break; 2271 } 2272 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL)) 2273 OuterExitL = NewOuterExitL; 2274 } 2275 2276 // At this point, we're definitely going to unswitch something so invalidate 2277 // any cached information in ScalarEvolution for the outer most loop 2278 // containing an exit block and all nested loops. 2279 if (SE) { 2280 if (OuterExitL) 2281 SE->forgetLoop(OuterExitL); 2282 else 2283 SE->forgetTopmostLoop(&L); 2284 SE->forgetBlockAndLoopDispositions(); 2285 } 2286 2287 // If the edge from this terminator to a successor dominates that successor, 2288 // store a map from each block in its dominator subtree to it. This lets us 2289 // tell when cloning for a particular successor if a block is dominated by 2290 // some *other* successor with a single data structure. We use this to 2291 // significantly reduce cloning. 2292 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc; 2293 for (auto *SuccBB : llvm::concat<BasicBlock *const>(ArrayRef(RetainedSuccBB), 2294 UnswitchedSuccBBs)) 2295 if (SuccBB->getUniquePredecessor() || 2296 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 2297 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB); 2298 })) 2299 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) { 2300 DominatingSucc[BB] = SuccBB; 2301 return true; 2302 }); 2303 2304 // Split the preheader, so that we know that there is a safe place to insert 2305 // the conditional branch. We will change the preheader to have a conditional 2306 // branch on LoopCond. The original preheader will become the split point 2307 // between the unswitched versions, and we will have a new preheader for the 2308 // original loop. 2309 BasicBlock *SplitBB = L.getLoopPreheader(); 2310 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU); 2311 2312 // Keep track of the dominator tree updates needed. 2313 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 2314 2315 // Clone the loop for each unswitched successor. 2316 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; 2317 VMaps.reserve(UnswitchedSuccBBs.size()); 2318 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs; 2319 for (auto *SuccBB : UnswitchedSuccBBs) { 2320 VMaps.emplace_back(new ValueToValueMapTy()); 2321 ClonedPHs[SuccBB] = buildClonedLoopBlocks( 2322 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB, 2323 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE); 2324 } 2325 2326 // Drop metadata if we may break its semantics by moving this instr into the 2327 // split block. 2328 if (TI.getMetadata(LLVMContext::MD_make_implicit)) { 2329 if (DropNonTrivialImplicitNullChecks) 2330 // Do not spend time trying to understand if we can keep it, just drop it 2331 // to save compile time. 2332 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr); 2333 else { 2334 // It is only legal to preserve make.implicit metadata if we are 2335 // guaranteed no reach implicit null check after following this branch. 2336 ICFLoopSafetyInfo SafetyInfo; 2337 SafetyInfo.computeLoopSafetyInfo(&L); 2338 if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L)) 2339 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr); 2340 } 2341 } 2342 2343 // The stitching of the branched code back together depends on whether we're 2344 // doing full unswitching or not with the exception that we always want to 2345 // nuke the initial terminator placed in the split block. 2346 SplitBB->getTerminator()->eraseFromParent(); 2347 if (FullUnswitch) { 2348 // Keep a clone of the terminator for MSSA updates. 2349 Instruction *NewTI = TI.clone(); 2350 NewTI->insertInto(ParentBB, ParentBB->end()); 2351 2352 // Splice the terminator from the original loop and rewrite its 2353 // successors. 2354 TI.moveBefore(*SplitBB, SplitBB->end()); 2355 TI.dropLocation(); 2356 2357 // First wire up the moved terminator to the preheaders. 2358 if (BI) { 2359 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2360 BI->setSuccessor(ClonedSucc, ClonedPH); 2361 BI->setSuccessor(1 - ClonedSucc, LoopPH); 2362 Value *Cond = skipTrivialSelect(BI->getCondition()); 2363 if (InsertFreeze) { 2364 // We don't give any debug location to the new freeze, because the 2365 // BI (`dyn_cast<BranchInst>(TI)`) is an in-loop instruction hoisted 2366 // out of the loop. 2367 Cond = new FreezeInst(Cond, Cond->getName() + ".fr", BI->getIterator()); 2368 } 2369 BI->setCondition(Cond); 2370 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2371 } else { 2372 assert(SI && "Must either be a branch or switch!"); 2373 2374 // Walk the cases and directly update their successors. 2375 assert(SI->getDefaultDest() == RetainedSuccBB && 2376 "Not retaining default successor!"); 2377 SI->setDefaultDest(LoopPH); 2378 for (const auto &Case : SI->cases()) 2379 if (Case.getCaseSuccessor() == RetainedSuccBB) 2380 Case.setSuccessor(LoopPH); 2381 else 2382 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second); 2383 2384 if (InsertFreeze) 2385 SI->setCondition(new FreezeInst(SI->getCondition(), 2386 SI->getCondition()->getName() + ".fr", 2387 SI->getIterator())); 2388 2389 // We need to use the set to populate domtree updates as even when there 2390 // are multiple cases pointing at the same successor we only want to 2391 // remove and insert one edge in the domtree. 2392 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2393 DTUpdates.push_back( 2394 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second}); 2395 } 2396 2397 if (MSSAU) { 2398 DT.applyUpdates(DTUpdates); 2399 DTUpdates.clear(); 2400 2401 // Remove all but one edge to the retained block and all unswitched 2402 // blocks. This is to avoid having duplicate entries in the cloned Phis, 2403 // when we know we only keep a single edge for each case. 2404 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB); 2405 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2406 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB); 2407 2408 for (auto &VMap : VMaps) 2409 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, 2410 /*IgnoreIncomingWithNoClones=*/true); 2411 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); 2412 2413 // Remove all edges to unswitched blocks. 2414 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2415 MSSAU->removeEdge(ParentBB, SuccBB); 2416 } 2417 2418 // Now unhook the successor relationship as we'll be replacing 2419 // the terminator with a direct branch. This is much simpler for branches 2420 // than switches so we handle those first. 2421 if (BI) { 2422 // Remove the parent as a predecessor of the unswitched successor. 2423 assert(UnswitchedSuccBBs.size() == 1 && 2424 "Only one possible unswitched block for a branch!"); 2425 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin(); 2426 UnswitchedSuccBB->removePredecessor(ParentBB, 2427 /*KeepOneInputPHIs*/ true); 2428 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB}); 2429 } else { 2430 // Note that we actually want to remove the parent block as a predecessor 2431 // of *every* case successor. The case successor is either unswitched, 2432 // completely eliminating an edge from the parent to that successor, or it 2433 // is a duplicate edge to the retained successor as the retained successor 2434 // is always the default successor and as we'll replace this with a direct 2435 // branch we no longer need the duplicate entries in the PHI nodes. 2436 SwitchInst *NewSI = cast<SwitchInst>(NewTI); 2437 assert(NewSI->getDefaultDest() == RetainedSuccBB && 2438 "Not retaining default successor!"); 2439 for (const auto &Case : NewSI->cases()) 2440 Case.getCaseSuccessor()->removePredecessor( 2441 ParentBB, 2442 /*KeepOneInputPHIs*/ true); 2443 2444 // We need to use the set to populate domtree updates as even when there 2445 // are multiple cases pointing at the same successor we only want to 2446 // remove and insert one edge in the domtree. 2447 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2448 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB}); 2449 } 2450 2451 // Create a new unconditional branch to the continuing block (as opposed to 2452 // the one cloned). 2453 Instruction *NewBI = BranchInst::Create(RetainedSuccBB, ParentBB); 2454 NewBI->setDebugLoc(NewTI->getDebugLoc()); 2455 2456 // After MSSAU update, remove the cloned terminator instruction NewTI. 2457 NewTI->eraseFromParent(); 2458 } else { 2459 assert(BI && "Only branches have partial unswitching."); 2460 assert(UnswitchedSuccBBs.size() == 1 && 2461 "Only one possible unswitched block for a branch!"); 2462 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2463 // When doing a partial unswitch, we have to do a bit more work to build up 2464 // the branch in the split block. 2465 if (PartiallyInvariant) 2466 buildPartialInvariantUnswitchConditionalBranch( 2467 *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU); 2468 else { 2469 buildPartialUnswitchConditionalBranch( 2470 *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, 2471 FreezeLoopUnswitchCond, BI, &AC, DT); 2472 } 2473 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2474 2475 if (MSSAU) { 2476 DT.applyUpdates(DTUpdates); 2477 DTUpdates.clear(); 2478 2479 // Perform MSSA cloning updates. 2480 for (auto &VMap : VMaps) 2481 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, 2482 /*IgnoreIncomingWithNoClones=*/true); 2483 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); 2484 } 2485 } 2486 2487 // Apply the updates accumulated above to get an up-to-date dominator tree. 2488 DT.applyUpdates(DTUpdates); 2489 2490 // Now that we have an accurate dominator tree, first delete the dead cloned 2491 // blocks so that we can accurately build any cloned loops. It is important to 2492 // not delete the blocks from the original loop yet because we still want to 2493 // reference the original loop to understand the cloned loop's structure. 2494 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU); 2495 2496 // Build the cloned loop structure itself. This may be substantially 2497 // different from the original structure due to the simplified CFG. This also 2498 // handles inserting all the cloned blocks into the correct loops. 2499 SmallVector<Loop *, 4> NonChildClonedLoops; 2500 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps) 2501 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops); 2502 2503 // Now that our cloned loops have been built, we can update the original loop. 2504 // First we delete the dead blocks from it and then we rebuild the loop 2505 // structure taking these deletions into account. 2506 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater); 2507 2508 if (MSSAU && VerifyMemorySSA) 2509 MSSAU->getMemorySSA()->verifyMemorySSA(); 2510 2511 SmallVector<Loop *, 4> HoistedLoops; 2512 bool IsStillLoop = 2513 rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE); 2514 2515 if (MSSAU && VerifyMemorySSA) 2516 MSSAU->getMemorySSA()->verifyMemorySSA(); 2517 2518 // This transformation has a high risk of corrupting the dominator tree, and 2519 // the below steps to rebuild loop structures will result in hard to debug 2520 // errors in that case so verify that the dominator tree is sane first. 2521 // FIXME: Remove this when the bugs stop showing up and rely on existing 2522 // verification steps. 2523 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2524 2525 if (BI && !PartiallyInvariant) { 2526 // If we unswitched a branch which collapses the condition to a known 2527 // constant we want to replace all the uses of the invariants within both 2528 // the original and cloned blocks. We do this here so that we can use the 2529 // now updated dominator tree to identify which side the users are on. 2530 assert(UnswitchedSuccBBs.size() == 1 && 2531 "Only one possible unswitched block for a branch!"); 2532 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2533 2534 // When considering multiple partially-unswitched invariants 2535 // we cant just go replace them with constants in both branches. 2536 // 2537 // For 'AND' we infer that true branch ("continue") means true 2538 // for each invariant operand. 2539 // For 'OR' we can infer that false branch ("continue") means false 2540 // for each invariant operand. 2541 // So it happens that for multiple-partial case we dont replace 2542 // in the unswitched branch. 2543 bool ReplaceUnswitched = 2544 FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant; 2545 2546 ConstantInt *UnswitchedReplacement = 2547 Direction ? ConstantInt::getTrue(BI->getContext()) 2548 : ConstantInt::getFalse(BI->getContext()); 2549 ConstantInt *ContinueReplacement = 2550 Direction ? ConstantInt::getFalse(BI->getContext()) 2551 : ConstantInt::getTrue(BI->getContext()); 2552 for (Value *Invariant : Invariants) { 2553 assert(!isa<Constant>(Invariant) && 2554 "Should not be replacing constant values!"); 2555 // Use make_early_inc_range here as set invalidates the iterator. 2556 for (Use &U : llvm::make_early_inc_range(Invariant->uses())) { 2557 Instruction *UserI = dyn_cast<Instruction>(U.getUser()); 2558 if (!UserI) 2559 continue; 2560 2561 // Replace it with the 'continue' side if in the main loop body, and the 2562 // unswitched if in the cloned blocks. 2563 if (DT.dominates(LoopPH, UserI->getParent())) 2564 U.set(ContinueReplacement); 2565 else if (ReplaceUnswitched && 2566 DT.dominates(ClonedPH, UserI->getParent())) 2567 U.set(UnswitchedReplacement); 2568 } 2569 } 2570 } 2571 2572 // We can change which blocks are exit blocks of all the cloned sibling 2573 // loops, the current loop, and any parent loops which shared exit blocks 2574 // with the current loop. As a consequence, we need to re-form LCSSA for 2575 // them. But we shouldn't need to re-form LCSSA for any child loops. 2576 // FIXME: This could be made more efficient by tracking which exit blocks are 2577 // new, and focusing on them, but that isn't likely to be necessary. 2578 // 2579 // In order to reasonably rebuild LCSSA we need to walk inside-out across the 2580 // loop nest and update every loop that could have had its exits changed. We 2581 // also need to cover any intervening loops. We add all of these loops to 2582 // a list and sort them by loop depth to achieve this without updating 2583 // unnecessary loops. 2584 auto UpdateLoop = [&](Loop &UpdateL) { 2585 #ifndef NDEBUG 2586 UpdateL.verifyLoop(); 2587 for (Loop *ChildL : UpdateL) { 2588 ChildL->verifyLoop(); 2589 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) && 2590 "Perturbed a child loop's LCSSA form!"); 2591 } 2592 #endif 2593 // First build LCSSA for this loop so that we can preserve it when 2594 // forming dedicated exits. We don't want to perturb some other loop's 2595 // LCSSA while doing that CFG edit. 2596 formLCSSA(UpdateL, DT, &LI, SE); 2597 2598 // For loops reached by this loop's original exit blocks we may 2599 // introduced new, non-dedicated exits. At least try to re-form dedicated 2600 // exits for these loops. This may fail if they couldn't have dedicated 2601 // exits to start with. 2602 formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true); 2603 }; 2604 2605 // For non-child cloned loops and hoisted loops, we just need to update LCSSA 2606 // and we can do it in any order as they don't nest relative to each other. 2607 // 2608 // Also check if any of the loops we have updated have become top-level loops 2609 // as that will necessitate widening the outer loop scope. 2610 for (Loop *UpdatedL : 2611 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) { 2612 UpdateLoop(*UpdatedL); 2613 if (UpdatedL->isOutermost()) 2614 OuterExitL = nullptr; 2615 } 2616 if (IsStillLoop) { 2617 UpdateLoop(L); 2618 if (L.isOutermost()) 2619 OuterExitL = nullptr; 2620 } 2621 2622 // If the original loop had exit blocks, walk up through the outer most loop 2623 // of those exit blocks to update LCSSA and form updated dedicated exits. 2624 if (OuterExitL != &L) 2625 for (Loop *OuterL = ParentL; OuterL != OuterExitL; 2626 OuterL = OuterL->getParentLoop()) 2627 UpdateLoop(*OuterL); 2628 2629 #ifndef NDEBUG 2630 // Verify the entire loop structure to catch any incorrect updates before we 2631 // progress in the pass pipeline. 2632 LI.verify(DT); 2633 #endif 2634 2635 // Now that we've unswitched something, make callbacks to report the changes. 2636 // For that we need to merge together the updated loops and the cloned loops 2637 // and check whether the original loop survived. 2638 SmallVector<Loop *, 4> SibLoops; 2639 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) 2640 if (UpdatedL->getParentLoop() == ParentL) 2641 SibLoops.push_back(UpdatedL); 2642 postUnswitch(L, LoopUpdater, LoopName, IsStillLoop, PartiallyInvariant, 2643 InjectedCondition, SibLoops); 2644 2645 if (MSSAU && VerifyMemorySSA) 2646 MSSAU->getMemorySSA()->verifyMemorySSA(); 2647 2648 if (BI) 2649 ++NumBranches; 2650 else 2651 ++NumSwitches; 2652 } 2653 2654 /// Recursively compute the cost of a dominator subtree based on the per-block 2655 /// cost map provided. 2656 /// 2657 /// The recursive computation is memozied into the provided DT-indexed cost map 2658 /// to allow querying it for most nodes in the domtree without it becoming 2659 /// quadratic. 2660 static InstructionCost computeDomSubtreeCost( 2661 DomTreeNode &N, 2662 const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap, 2663 SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) { 2664 // Don't accumulate cost (or recurse through) blocks not in our block cost 2665 // map and thus not part of the duplication cost being considered. 2666 auto BBCostIt = BBCostMap.find(N.getBlock()); 2667 if (BBCostIt == BBCostMap.end()) 2668 return 0; 2669 2670 // Lookup this node to see if we already computed its cost. 2671 auto DTCostIt = DTCostMap.find(&N); 2672 if (DTCostIt != DTCostMap.end()) 2673 return DTCostIt->second; 2674 2675 // If not, we have to compute it. We can't use insert above and update 2676 // because computing the cost may insert more things into the map. 2677 InstructionCost Cost = std::accumulate( 2678 N.begin(), N.end(), BBCostIt->second, 2679 [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost { 2680 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap); 2681 }); 2682 bool Inserted = DTCostMap.insert({&N, Cost}).second; 2683 (void)Inserted; 2684 assert(Inserted && "Should not insert a node while visiting children!"); 2685 return Cost; 2686 } 2687 2688 /// Turns a select instruction into implicit control flow branch, 2689 /// making the following replacement: 2690 /// 2691 /// head: 2692 /// --code before select-- 2693 /// select %cond, %trueval, %falseval 2694 /// --code after select-- 2695 /// 2696 /// into 2697 /// 2698 /// head: 2699 /// --code before select-- 2700 /// br i1 %cond, label %then, label %tail 2701 /// 2702 /// then: 2703 /// br %tail 2704 /// 2705 /// tail: 2706 /// phi [ %trueval, %then ], [ %falseval, %head] 2707 /// unreachable 2708 /// 2709 /// It also makes all relevant DT and LI updates, so that all structures are in 2710 /// valid state after this transform. 2711 static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT, 2712 LoopInfo &LI, MemorySSAUpdater *MSSAU, 2713 AssumptionCache *AC) { 2714 LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n"); 2715 BasicBlock *HeadBB = SI->getParent(); 2716 2717 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 2718 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false, 2719 SI->getMetadata(LLVMContext::MD_prof), &DTU, &LI); 2720 auto *CondBr = cast<BranchInst>(HeadBB->getTerminator()); 2721 BasicBlock *ThenBB = CondBr->getSuccessor(0), 2722 *TailBB = CondBr->getSuccessor(1); 2723 if (MSSAU) 2724 MSSAU->moveAllAfterSpliceBlocks(HeadBB, TailBB, SI); 2725 2726 PHINode *Phi = 2727 PHINode::Create(SI->getType(), 2, "unswitched.select", SI->getIterator()); 2728 Phi->addIncoming(SI->getTrueValue(), ThenBB); 2729 Phi->addIncoming(SI->getFalseValue(), HeadBB); 2730 Phi->setDebugLoc(SI->getDebugLoc()); 2731 SI->replaceAllUsesWith(Phi); 2732 SI->eraseFromParent(); 2733 2734 if (MSSAU && VerifyMemorySSA) 2735 MSSAU->getMemorySSA()->verifyMemorySSA(); 2736 2737 ++NumSelects; 2738 return CondBr; 2739 } 2740 2741 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch, 2742 /// making the following replacement: 2743 /// 2744 /// --code before guard-- 2745 /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ] 2746 /// --code after guard-- 2747 /// 2748 /// into 2749 /// 2750 /// --code before guard-- 2751 /// br i1 %cond, label %guarded, label %deopt 2752 /// 2753 /// guarded: 2754 /// --code after guard-- 2755 /// 2756 /// deopt: 2757 /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ] 2758 /// unreachable 2759 /// 2760 /// It also makes all relevant DT and LI updates, so that all structures are in 2761 /// valid state after this transform. 2762 static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L, 2763 DominatorTree &DT, LoopInfo &LI, 2764 MemorySSAUpdater *MSSAU) { 2765 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 2766 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n"); 2767 BasicBlock *CheckBB = GI->getParent(); 2768 2769 if (MSSAU && VerifyMemorySSA) 2770 MSSAU->getMemorySSA()->verifyMemorySSA(); 2771 2772 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 2773 Instruction *DeoptBlockTerm = 2774 SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true, 2775 GI->getMetadata(LLVMContext::MD_prof), &DTU, &LI); 2776 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator()); 2777 // SplitBlockAndInsertIfThen inserts control flow that branches to 2778 // DeoptBlockTerm if the condition is true. We want the opposite. 2779 CheckBI->swapSuccessors(); 2780 2781 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0); 2782 GuardedBlock->setName("guarded"); 2783 CheckBI->getSuccessor(1)->setName("deopt"); 2784 BasicBlock *DeoptBlock = CheckBI->getSuccessor(1); 2785 2786 if (MSSAU) 2787 MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI); 2788 2789 GI->moveBefore(DeoptBlockTerm); 2790 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext())); 2791 2792 if (MSSAU) { 2793 MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI)); 2794 MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator); 2795 if (VerifyMemorySSA) 2796 MSSAU->getMemorySSA()->verifyMemorySSA(); 2797 } 2798 2799 if (VerifyLoopInfo) 2800 LI.verify(DT); 2801 ++NumGuards; 2802 return CheckBI; 2803 } 2804 2805 /// Cost multiplier is a way to limit potentially exponential behavior 2806 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch 2807 /// candidates available. Also accounting for the number of "sibling" loops with 2808 /// the idea to account for previous unswitches that already happened on this 2809 /// cluster of loops. There was an attempt to keep this formula simple, 2810 /// just enough to limit the worst case behavior. Even if it is not that simple 2811 /// now it is still not an attempt to provide a detailed heuristic size 2812 /// prediction. 2813 /// 2814 /// TODO: Make a proper accounting of "explosion" effect for all kinds of 2815 /// unswitch candidates, making adequate predictions instead of wild guesses. 2816 /// That requires knowing not just the number of "remaining" candidates but 2817 /// also costs of unswitching for each of these candidates. 2818 static int CalculateUnswitchCostMultiplier( 2819 const Instruction &TI, const Loop &L, const LoopInfo &LI, 2820 const DominatorTree &DT, 2821 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) { 2822 2823 // Guards and other exiting conditions do not contribute to exponential 2824 // explosion as soon as they dominate the latch (otherwise there might be 2825 // another path to the latch remaining that does not allow to eliminate the 2826 // loop copy on unswitch). 2827 const BasicBlock *Latch = L.getLoopLatch(); 2828 const BasicBlock *CondBlock = TI.getParent(); 2829 if (DT.dominates(CondBlock, Latch) && 2830 (isGuard(&TI) || 2831 (TI.isTerminator() && 2832 llvm::count_if(successors(&TI), [&L](const BasicBlock *SuccBB) { 2833 return L.contains(SuccBB); 2834 }) <= 1))) { 2835 NumCostMultiplierSkipped++; 2836 return 1; 2837 } 2838 2839 auto *ParentL = L.getParentLoop(); 2840 int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size() 2841 : std::distance(LI.begin(), LI.end())); 2842 // Count amount of clones that all the candidates might cause during 2843 // unswitching. Branch/guard/select counts as 1, switch counts as log2 of its 2844 // cases. 2845 int UnswitchedClones = 0; 2846 for (const auto &Candidate : UnswitchCandidates) { 2847 const Instruction *CI = Candidate.TI; 2848 const BasicBlock *CondBlock = CI->getParent(); 2849 bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch); 2850 if (isa<SelectInst>(CI)) { 2851 UnswitchedClones++; 2852 continue; 2853 } 2854 if (isGuard(CI)) { 2855 if (!SkipExitingSuccessors) 2856 UnswitchedClones++; 2857 continue; 2858 } 2859 int NonExitingSuccessors = 2860 llvm::count_if(successors(CondBlock), 2861 [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) { 2862 return !SkipExitingSuccessors || L.contains(SuccBB); 2863 }); 2864 UnswitchedClones += Log2_32(NonExitingSuccessors); 2865 } 2866 2867 // Ignore up to the "unscaled candidates" number of unswitch candidates 2868 // when calculating the power-of-two scaling of the cost. The main idea 2869 // with this control is to allow a small number of unswitches to happen 2870 // and rely more on siblings multiplier (see below) when the number 2871 // of candidates is small. 2872 unsigned ClonesPower = 2873 std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0); 2874 2875 // Allowing top-level loops to spread a bit more than nested ones. 2876 int SiblingsMultiplier = 2877 std::max((ParentL ? SiblingsCount 2878 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv), 2879 1); 2880 // Compute the cost multiplier in a way that won't overflow by saturating 2881 // at an upper bound. 2882 int CostMultiplier; 2883 if (ClonesPower > Log2_32(UnswitchThreshold) || 2884 SiblingsMultiplier > UnswitchThreshold) 2885 CostMultiplier = UnswitchThreshold; 2886 else 2887 CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower), 2888 (int)UnswitchThreshold); 2889 2890 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier 2891 << " (siblings " << SiblingsMultiplier << " * clones " 2892 << (1 << ClonesPower) << ")" 2893 << " for unswitch candidate: " << TI << "\n"); 2894 return CostMultiplier; 2895 } 2896 2897 static bool collectUnswitchCandidates( 2898 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, 2899 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, 2900 const Loop &L, const LoopInfo &LI, AAResults &AA, 2901 const MemorySSAUpdater *MSSAU) { 2902 assert(UnswitchCandidates.empty() && "Should be!"); 2903 2904 auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) { 2905 Cond = skipTrivialSelect(Cond); 2906 if (isa<Constant>(Cond)) 2907 return; 2908 if (L.isLoopInvariant(Cond)) { 2909 UnswitchCandidates.push_back({I, {Cond}}); 2910 return; 2911 } 2912 if (match(Cond, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) { 2913 TinyPtrVector<Value *> Invariants = 2914 collectHomogenousInstGraphLoopInvariants( 2915 L, *static_cast<Instruction *>(Cond), LI); 2916 if (!Invariants.empty()) 2917 UnswitchCandidates.push_back({I, std::move(Invariants)}); 2918 } 2919 }; 2920 2921 // Whether or not we should also collect guards in the loop. 2922 bool CollectGuards = false; 2923 if (UnswitchGuards) { 2924 auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction( 2925 Intrinsic::getName(Intrinsic::experimental_guard)); 2926 if (GuardDecl && !GuardDecl->use_empty()) 2927 CollectGuards = true; 2928 } 2929 2930 for (auto *BB : L.blocks()) { 2931 if (LI.getLoopFor(BB) != &L) 2932 continue; 2933 2934 for (auto &I : *BB) { 2935 if (auto *SI = dyn_cast<SelectInst>(&I)) { 2936 auto *Cond = SI->getCondition(); 2937 // Do not unswitch vector selects and logical and/or selects 2938 if (Cond->getType()->isIntegerTy(1) && !SI->getType()->isIntegerTy(1)) 2939 AddUnswitchCandidatesForInst(SI, Cond); 2940 } else if (CollectGuards && isGuard(&I)) { 2941 auto *Cond = 2942 skipTrivialSelect(cast<IntrinsicInst>(&I)->getArgOperand(0)); 2943 // TODO: Support AND, OR conditions and partial unswitching. 2944 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond)) 2945 UnswitchCandidates.push_back({&I, {Cond}}); 2946 } 2947 } 2948 2949 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2950 // We can only consider fully loop-invariant switch conditions as we need 2951 // to completely eliminate the switch after unswitching. 2952 if (!isa<Constant>(SI->getCondition()) && 2953 L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor()) 2954 UnswitchCandidates.push_back({SI, {SI->getCondition()}}); 2955 continue; 2956 } 2957 2958 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 2959 if (!BI || !BI->isConditional() || 2960 BI->getSuccessor(0) == BI->getSuccessor(1)) 2961 continue; 2962 2963 AddUnswitchCandidatesForInst(BI, BI->getCondition()); 2964 } 2965 2966 if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") && 2967 !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) { 2968 return TerminatorAndInvariants.TI == L.getHeader()->getTerminator(); 2969 })) { 2970 MemorySSA *MSSA = MSSAU->getMemorySSA(); 2971 if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) { 2972 LLVM_DEBUG( 2973 dbgs() << "simple-loop-unswitch: Found partially invariant condition " 2974 << *Info->InstToDuplicate[0] << "\n"); 2975 PartialIVInfo = *Info; 2976 PartialIVCondBranch = L.getHeader()->getTerminator(); 2977 TinyPtrVector<Value *> ValsToDuplicate; 2978 llvm::append_range(ValsToDuplicate, Info->InstToDuplicate); 2979 UnswitchCandidates.push_back( 2980 {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)}); 2981 } 2982 } 2983 return !UnswitchCandidates.empty(); 2984 } 2985 2986 /// Tries to canonicalize condition described by: 2987 /// 2988 /// br (LHS pred RHS), label IfTrue, label IfFalse 2989 /// 2990 /// into its equivalent where `Pred` is something that we support for injected 2991 /// invariants (so far it is limited to ult), LHS in canonicalized form is 2992 /// non-invariant and RHS is an invariant. 2993 static void canonicalizeForInvariantConditionInjection( 2994 ICmpInst::Predicate &Pred, Value *&LHS, Value *&RHS, BasicBlock *&IfTrue, 2995 BasicBlock *&IfFalse, const Loop &L) { 2996 if (!L.contains(IfTrue)) { 2997 Pred = ICmpInst::getInversePredicate(Pred); 2998 std::swap(IfTrue, IfFalse); 2999 } 3000 3001 // Move loop-invariant argument to RHS position. 3002 if (L.isLoopInvariant(LHS)) { 3003 Pred = ICmpInst::getSwappedPredicate(Pred); 3004 std::swap(LHS, RHS); 3005 } 3006 3007 if (Pred == ICmpInst::ICMP_SGE && match(RHS, m_Zero())) { 3008 // Turn "x >=s 0" into "x <u UMIN_INT" 3009 Pred = ICmpInst::ICMP_ULT; 3010 RHS = ConstantInt::get( 3011 RHS->getContext(), 3012 APInt::getSignedMinValue(RHS->getType()->getIntegerBitWidth())); 3013 } 3014 } 3015 3016 /// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS ) 3017 /// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by 3018 /// injecting a loop-invariant condition. 3019 static bool shouldTryInjectInvariantCondition( 3020 const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS, 3021 const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) { 3022 if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS)) 3023 return false; 3024 // TODO: Support other predicates. 3025 if (Pred != ICmpInst::ICMP_ULT) 3026 return false; 3027 // TODO: Support non-loop-exiting branches? 3028 if (!L.contains(IfTrue) || L.contains(IfFalse)) 3029 return false; 3030 // FIXME: For some reason this causes problems with MSSA updates, need to 3031 // investigate why. So far, just don't unswitch latch. 3032 if (L.getHeader() == IfTrue) 3033 return false; 3034 return true; 3035 } 3036 3037 /// Returns true, if metadata on \p BI allows us to optimize branching into \p 3038 /// TakenSucc via injection of invariant conditions. The branch should be not 3039 /// enough and not previously unswitched, the information about this comes from 3040 /// the metadata. 3041 bool shouldTryInjectBasingOnMetadata(const BranchInst *BI, 3042 const BasicBlock *TakenSucc) { 3043 SmallVector<uint32_t> Weights; 3044 if (!extractBranchWeights(*BI, Weights)) 3045 return false; 3046 unsigned T = InjectInvariantConditionHotnesThreshold; 3047 BranchProbability LikelyTaken(T - 1, T); 3048 3049 assert(Weights.size() == 2 && "Unexpected profile data!"); 3050 size_t Idx = BI->getSuccessor(0) == TakenSucc ? 0 : 1; 3051 auto Num = Weights[Idx]; 3052 auto Denom = Weights[0] + Weights[1]; 3053 // Degenerate or overflowed metadata. 3054 if (Denom == 0 || Num > Denom) 3055 return false; 3056 BranchProbability ActualTaken(Num, Denom); 3057 if (LikelyTaken > ActualTaken) 3058 return false; 3059 return true; 3060 } 3061 3062 /// Materialize pending invariant condition of the given candidate into IR. The 3063 /// injected loop-invariant condition implies the original loop-variant branch 3064 /// condition, so the materialization turns 3065 /// 3066 /// loop_block: 3067 /// ... 3068 /// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc 3069 /// 3070 /// into 3071 /// 3072 /// preheader: 3073 /// %invariant_cond = LHS pred RHS 3074 /// ... 3075 /// loop_block: 3076 /// br i1 %invariant_cond, label InLoopSucc, label OriginalCheck 3077 /// OriginalCheck: 3078 /// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc 3079 /// ... 3080 static NonTrivialUnswitchCandidate 3081 injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L, 3082 DominatorTree &DT, LoopInfo &LI, 3083 AssumptionCache &AC, MemorySSAUpdater *MSSAU) { 3084 assert(Candidate.hasPendingInjection() && "Nothing to inject!"); 3085 BasicBlock *Preheader = L.getLoopPreheader(); 3086 assert(Preheader && "Loop is not in simplified form?"); 3087 assert(LI.getLoopFor(Candidate.TI->getParent()) == &L && 3088 "Unswitching branch of inner loop!"); 3089 3090 auto Pred = Candidate.PendingInjection->Pred; 3091 auto *LHS = Candidate.PendingInjection->LHS; 3092 auto *RHS = Candidate.PendingInjection->RHS; 3093 auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc; 3094 auto *TI = cast<BranchInst>(Candidate.TI); 3095 auto *BB = Candidate.TI->getParent(); 3096 auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(0) ? TI->getSuccessor(1) 3097 : TI->getSuccessor(0); 3098 // FIXME: Remove this once limitation on successors is lifted. 3099 assert(L.contains(InLoopSucc) && "Not supported yet!"); 3100 assert(!L.contains(OutOfLoopSucc) && "Not supported yet!"); 3101 auto &Ctx = BB->getContext(); 3102 3103 IRBuilder<> Builder(Preheader->getTerminator()); 3104 assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!"); 3105 if (LHS->getType() != RHS->getType()) { 3106 if (LHS->getType()->getIntegerBitWidth() < 3107 RHS->getType()->getIntegerBitWidth()) 3108 LHS = Builder.CreateZExt(LHS, RHS->getType(), LHS->getName() + ".wide"); 3109 else 3110 RHS = Builder.CreateZExt(RHS, LHS->getType(), RHS->getName() + ".wide"); 3111 } 3112 // Do not use builder here: CreateICmp may simplify this into a constant and 3113 // unswitching will break. Better optimize it away later. 3114 auto *InjectedCond = 3115 ICmpInst::Create(Instruction::ICmp, Pred, LHS, RHS, "injected.cond", 3116 Preheader->getTerminator()->getIterator()); 3117 3118 BasicBlock *CheckBlock = BasicBlock::Create(Ctx, BB->getName() + ".check", 3119 BB->getParent(), InLoopSucc); 3120 Builder.SetInsertPoint(TI); 3121 auto *InvariantBr = 3122 Builder.CreateCondBr(InjectedCond, InLoopSucc, CheckBlock); 3123 3124 Builder.SetInsertPoint(CheckBlock); 3125 Builder.CreateCondBr(TI->getCondition(), TI->getSuccessor(0), 3126 TI->getSuccessor(1)); 3127 TI->eraseFromParent(); 3128 3129 // Fixup phis. 3130 for (auto &I : *InLoopSucc) { 3131 auto *PN = dyn_cast<PHINode>(&I); 3132 if (!PN) 3133 break; 3134 auto *Inc = PN->getIncomingValueForBlock(BB); 3135 PN->addIncoming(Inc, CheckBlock); 3136 } 3137 OutOfLoopSucc->replacePhiUsesWith(BB, CheckBlock); 3138 3139 SmallVector<DominatorTree::UpdateType, 4> DTUpdates = { 3140 { DominatorTree::Insert, BB, CheckBlock }, 3141 { DominatorTree::Insert, CheckBlock, InLoopSucc }, 3142 { DominatorTree::Insert, CheckBlock, OutOfLoopSucc }, 3143 { DominatorTree::Delete, BB, OutOfLoopSucc } 3144 }; 3145 3146 DT.applyUpdates(DTUpdates); 3147 if (MSSAU) 3148 MSSAU->applyUpdates(DTUpdates, DT); 3149 L.addBasicBlockToLoop(CheckBlock, LI); 3150 3151 #ifndef NDEBUG 3152 DT.verify(); 3153 LI.verify(DT); 3154 if (MSSAU && VerifyMemorySSA) 3155 MSSAU->getMemorySSA()->verifyMemorySSA(); 3156 #endif 3157 3158 // TODO: In fact, cost of unswitching a new invariant candidate is *slightly* 3159 // higher because we have just inserted a new block. Need to think how to 3160 // adjust the cost of injected candidates when it was first computed. 3161 LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr 3162 << " and considering it for unswitching."); 3163 ++NumInvariantConditionsInjected; 3164 return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond }, 3165 Candidate.Cost); 3166 } 3167 3168 /// Given chain of loop branch conditions looking like: 3169 /// br (Variant < Invariant1) 3170 /// br (Variant < Invariant2) 3171 /// br (Variant < Invariant3) 3172 /// ... 3173 /// collect set of invariant conditions on which we want to unswitch, which 3174 /// look like: 3175 /// Invariant1 <= Invariant2 3176 /// Invariant2 <= Invariant3 3177 /// ... 3178 /// Though they might not immediately exist in the IR, we can still inject them. 3179 static bool insertCandidatesWithPendingInjections( 3180 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L, 3181 ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares, 3182 const DominatorTree &DT) { 3183 3184 assert(ICmpInst::isRelational(Pred)); 3185 assert(ICmpInst::isStrictPredicate(Pred)); 3186 if (Compares.size() < 2) 3187 return false; 3188 ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(Pred); 3189 for (auto Prev = Compares.begin(), Next = Compares.begin() + 1; 3190 Next != Compares.end(); ++Prev, ++Next) { 3191 Value *LHS = Next->Invariant; 3192 Value *RHS = Prev->Invariant; 3193 BasicBlock *InLoopSucc = Prev->InLoopSucc; 3194 InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc); 3195 NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS }, 3196 std::nullopt, std::move(ToInject)); 3197 UnswitchCandidates.push_back(std::move(Candidate)); 3198 } 3199 return true; 3200 } 3201 3202 /// Collect unswitch candidates by invariant conditions that are not immediately 3203 /// present in the loop. However, they can be injected into the code if we 3204 /// decide it's profitable. 3205 /// An example of such conditions is following: 3206 /// 3207 /// for (...) { 3208 /// x = load ... 3209 /// if (! x <u C1) break; 3210 /// if (! x <u C2) break; 3211 /// <do something> 3212 /// } 3213 /// 3214 /// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <= 3215 /// C2" automatically implies "x <u C2", so we can get rid of one of 3216 /// loop-variant checks in unswitched loop version. 3217 static bool collectUnswitchCandidatesWithInjections( 3218 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, 3219 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L, 3220 const DominatorTree &DT, const LoopInfo &LI, AAResults &AA, 3221 const MemorySSAUpdater *MSSAU) { 3222 if (!InjectInvariantConditions) 3223 return false; 3224 3225 if (!DT.isReachableFromEntry(L.getHeader())) 3226 return false; 3227 auto *Latch = L.getLoopLatch(); 3228 // Need to have a single latch and a preheader. 3229 if (!Latch) 3230 return false; 3231 assert(L.getLoopPreheader() && "Must have a preheader!"); 3232 3233 DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT; 3234 // Traverse the conditions that dominate latch (and therefore dominate each 3235 // other). 3236 for (auto *DTN = DT.getNode(Latch); L.contains(DTN->getBlock()); 3237 DTN = DTN->getIDom()) { 3238 ICmpInst::Predicate Pred; 3239 Value *LHS = nullptr, *RHS = nullptr; 3240 BasicBlock *IfTrue = nullptr, *IfFalse = nullptr; 3241 auto *BB = DTN->getBlock(); 3242 // Ignore inner loops. 3243 if (LI.getLoopFor(BB) != &L) 3244 continue; 3245 auto *Term = BB->getTerminator(); 3246 if (!match(Term, m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)), 3247 m_BasicBlock(IfTrue), m_BasicBlock(IfFalse)))) 3248 continue; 3249 if (!LHS->getType()->isIntegerTy()) 3250 continue; 3251 canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse, 3252 L); 3253 if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L)) 3254 continue; 3255 if (!shouldTryInjectBasingOnMetadata(cast<BranchInst>(Term), IfTrue)) 3256 continue; 3257 // Strip ZEXT for unsigned predicate. 3258 // TODO: once signed predicates are supported, also strip SEXT. 3259 CompareDesc Desc(cast<BranchInst>(Term), RHS, IfTrue); 3260 while (auto *Zext = dyn_cast<ZExtInst>(LHS)) 3261 LHS = Zext->getOperand(0); 3262 CandidatesULT[LHS].push_back(Desc); 3263 } 3264 3265 bool Found = false; 3266 for (auto &It : CandidatesULT) 3267 Found |= insertCandidatesWithPendingInjections( 3268 UnswitchCandidates, L, ICmpInst::ICMP_ULT, It.second, DT); 3269 return Found; 3270 } 3271 3272 static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) { 3273 if (!L.isSafeToClone()) 3274 return false; 3275 for (auto *BB : L.blocks()) 3276 for (auto &I : *BB) { 3277 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB)) 3278 return false; 3279 if (auto *CB = dyn_cast<CallBase>(&I)) { 3280 assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone()."); 3281 if (CB->isConvergent()) 3282 return false; 3283 } 3284 } 3285 3286 // Check if there are irreducible CFG cycles in this loop. If so, we cannot 3287 // easily unswitch non-trivial edges out of the loop. Doing so might turn the 3288 // irreducible control flow into reducible control flow and introduce new 3289 // loops "out of thin air". If we ever discover important use cases for doing 3290 // this, we can add support to loop unswitch, but it is a lot of complexity 3291 // for what seems little or no real world benefit. 3292 LoopBlocksRPO RPOT(&L); 3293 RPOT.perform(&LI); 3294 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 3295 return false; 3296 3297 SmallVector<BasicBlock *, 4> ExitBlocks; 3298 L.getUniqueExitBlocks(ExitBlocks); 3299 // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch 3300 // instruction as we don't know how to split those exit blocks. 3301 // FIXME: We should teach SplitBlock to handle this and remove this 3302 // restriction. 3303 for (auto *ExitBB : ExitBlocks) { 3304 auto *I = ExitBB->getFirstNonPHI(); 3305 if (isa<CleanupPadInst>(I) || isa<CatchSwitchInst>(I)) { 3306 LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch " 3307 "in exit block\n"); 3308 return false; 3309 } 3310 } 3311 3312 return true; 3313 } 3314 3315 static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate( 3316 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L, 3317 const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC, 3318 const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) { 3319 // Given that unswitching these terminators will require duplicating parts of 3320 // the loop, so we need to be able to model that cost. Compute the ephemeral 3321 // values and set up a data structure to hold per-BB costs. We cache each 3322 // block's cost so that we don't recompute this when considering different 3323 // subsets of the loop for duplication during unswitching. 3324 SmallPtrSet<const Value *, 4> EphValues; 3325 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues); 3326 SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap; 3327 3328 // Compute the cost of each block, as well as the total loop cost. Also, bail 3329 // out if we see instructions which are incompatible with loop unswitching 3330 // (convergent, noduplicate, or cross-basic-block tokens). 3331 // FIXME: We might be able to safely handle some of these in non-duplicated 3332 // regions. 3333 TargetTransformInfo::TargetCostKind CostKind = 3334 L.getHeader()->getParent()->hasMinSize() 3335 ? TargetTransformInfo::TCK_CodeSize 3336 : TargetTransformInfo::TCK_SizeAndLatency; 3337 InstructionCost LoopCost = 0; 3338 for (auto *BB : L.blocks()) { 3339 InstructionCost Cost = 0; 3340 for (auto &I : *BB) { 3341 if (EphValues.count(&I)) 3342 continue; 3343 Cost += TTI.getInstructionCost(&I, CostKind); 3344 } 3345 assert(Cost >= 0 && "Must not have negative costs!"); 3346 LoopCost += Cost; 3347 assert(LoopCost >= 0 && "Must not have negative loop costs!"); 3348 BBCostMap[BB] = Cost; 3349 } 3350 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n"); 3351 3352 // Now we find the best candidate by searching for the one with the following 3353 // properties in order: 3354 // 3355 // 1) An unswitching cost below the threshold 3356 // 2) The smallest number of duplicated unswitch candidates (to avoid 3357 // creating redundant subsequent unswitching) 3358 // 3) The smallest cost after unswitching. 3359 // 3360 // We prioritize reducing fanout of unswitch candidates provided the cost 3361 // remains below the threshold because this has a multiplicative effect. 3362 // 3363 // This requires memoizing each dominator subtree to avoid redundant work. 3364 // 3365 // FIXME: Need to actually do the number of candidates part above. 3366 SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap; 3367 // Given a terminator which might be unswitched, computes the non-duplicated 3368 // cost for that terminator. 3369 auto ComputeUnswitchedCost = [&](Instruction &TI, 3370 bool FullUnswitch) -> InstructionCost { 3371 // Unswitching selects unswitches the entire loop. 3372 if (isa<SelectInst>(TI)) 3373 return LoopCost; 3374 3375 BasicBlock &BB = *TI.getParent(); 3376 SmallPtrSet<BasicBlock *, 4> Visited; 3377 3378 InstructionCost Cost = 0; 3379 for (BasicBlock *SuccBB : successors(&BB)) { 3380 // Don't count successors more than once. 3381 if (!Visited.insert(SuccBB).second) 3382 continue; 3383 3384 // If this is a partial unswitch candidate, then it must be a conditional 3385 // branch with a condition of either `or`, `and`, their corresponding 3386 // select forms or partially invariant instructions. In that case, one of 3387 // the successors is necessarily duplicated, so don't even try to remove 3388 // its cost. 3389 if (!FullUnswitch) { 3390 auto &BI = cast<BranchInst>(TI); 3391 Value *Cond = skipTrivialSelect(BI.getCondition()); 3392 if (match(Cond, m_LogicalAnd())) { 3393 if (SuccBB == BI.getSuccessor(1)) 3394 continue; 3395 } else if (match(Cond, m_LogicalOr())) { 3396 if (SuccBB == BI.getSuccessor(0)) 3397 continue; 3398 } else if ((PartialIVInfo.KnownValue->isOneValue() && 3399 SuccBB == BI.getSuccessor(0)) || 3400 (!PartialIVInfo.KnownValue->isOneValue() && 3401 SuccBB == BI.getSuccessor(1))) 3402 continue; 3403 } 3404 3405 // This successor's domtree will not need to be duplicated after 3406 // unswitching if the edge to the successor dominates it (and thus the 3407 // entire tree). This essentially means there is no other path into this 3408 // subtree and so it will end up live in only one clone of the loop. 3409 if (SuccBB->getUniquePredecessor() || 3410 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 3411 return PredBB == &BB || DT.dominates(SuccBB, PredBB); 3412 })) { 3413 Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap); 3414 assert(Cost <= LoopCost && 3415 "Non-duplicated cost should never exceed total loop cost!"); 3416 } 3417 } 3418 3419 // Now scale the cost by the number of unique successors minus one. We 3420 // subtract one because there is already at least one copy of the entire 3421 // loop. This is computing the new cost of unswitching a condition. 3422 // Note that guards always have 2 unique successors that are implicit and 3423 // will be materialized if we decide to unswitch it. 3424 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size(); 3425 assert(SuccessorsCount > 1 && 3426 "Cannot unswitch a condition without multiple distinct successors!"); 3427 return (LoopCost - Cost) * (SuccessorsCount - 1); 3428 }; 3429 3430 std::optional<NonTrivialUnswitchCandidate> Best; 3431 for (auto &Candidate : UnswitchCandidates) { 3432 Instruction &TI = *Candidate.TI; 3433 ArrayRef<Value *> Invariants = Candidate.Invariants; 3434 BranchInst *BI = dyn_cast<BranchInst>(&TI); 3435 bool FullUnswitch = 3436 !BI || Candidate.hasPendingInjection() || 3437 (Invariants.size() == 1 && 3438 Invariants[0] == skipTrivialSelect(BI->getCondition())); 3439 InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch); 3440 // Calculate cost multiplier which is a tool to limit potentially 3441 // exponential behavior of loop-unswitch. 3442 if (EnableUnswitchCostMultiplier) { 3443 int CostMultiplier = 3444 CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates); 3445 assert( 3446 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) && 3447 "cost multiplier needs to be in the range of 1..UnswitchThreshold"); 3448 CandidateCost *= CostMultiplier; 3449 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 3450 << " (multiplier: " << CostMultiplier << ")" 3451 << " for unswitch candidate: " << TI << "\n"); 3452 } else { 3453 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 3454 << " for unswitch candidate: " << TI << "\n"); 3455 } 3456 3457 if (!Best || CandidateCost < Best->Cost) { 3458 Best = Candidate; 3459 Best->Cost = CandidateCost; 3460 } 3461 } 3462 assert(Best && "Must be!"); 3463 return *Best; 3464 } 3465 3466 // Insert a freeze on an unswitched branch if all is true: 3467 // 1. freeze-loop-unswitch-cond option is true 3468 // 2. The branch may not execute in the loop pre-transformation. If a branch may 3469 // not execute and could cause UB, it would always cause UB if it is hoisted outside 3470 // of the loop. Insert a freeze to prevent this case. 3471 // 3. The branch condition may be poison or undef 3472 static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT, 3473 AssumptionCache &AC) { 3474 assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI)); 3475 if (!FreezeLoopUnswitchCond) 3476 return false; 3477 3478 ICFLoopSafetyInfo SafetyInfo; 3479 SafetyInfo.computeLoopSafetyInfo(&L); 3480 if (SafetyInfo.isGuaranteedToExecute(TI, &DT, &L)) 3481 return false; 3482 3483 Value *Cond; 3484 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) 3485 Cond = skipTrivialSelect(BI->getCondition()); 3486 else 3487 Cond = skipTrivialSelect(cast<SwitchInst>(&TI)->getCondition()); 3488 return !isGuaranteedNotToBeUndefOrPoison( 3489 Cond, &AC, L.getLoopPreheader()->getTerminator(), &DT); 3490 } 3491 3492 static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI, 3493 AssumptionCache &AC, AAResults &AA, 3494 TargetTransformInfo &TTI, ScalarEvolution *SE, 3495 MemorySSAUpdater *MSSAU, 3496 LPMUpdater &LoopUpdater) { 3497 // Collect all invariant conditions within this loop (as opposed to an inner 3498 // loop which would be handled when visiting that inner loop). 3499 SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates; 3500 IVConditionInfo PartialIVInfo; 3501 Instruction *PartialIVCondBranch = nullptr; 3502 collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo, 3503 PartialIVCondBranch, L, LI, AA, MSSAU); 3504 if (!findOptionMDForLoop(&L, "llvm.loop.unswitch.injection.disable")) 3505 collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo, 3506 PartialIVCondBranch, L, DT, LI, AA, 3507 MSSAU); 3508 // If we didn't find any candidates, we're done. 3509 if (UnswitchCandidates.empty()) 3510 return false; 3511 3512 LLVM_DEBUG( 3513 dbgs() << "Considering " << UnswitchCandidates.size() 3514 << " non-trivial loop invariant conditions for unswitching.\n"); 3515 3516 NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate( 3517 UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo); 3518 3519 assert(Best.TI && "Failed to find loop unswitch candidate"); 3520 assert(Best.Cost && "Failed to compute cost"); 3521 3522 if (*Best.Cost >= UnswitchThreshold) { 3523 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost 3524 << "\n"); 3525 return false; 3526 } 3527 3528 bool InjectedCondition = false; 3529 if (Best.hasPendingInjection()) { 3530 Best = injectPendingInvariantConditions(Best, L, DT, LI, AC, MSSAU); 3531 InjectedCondition = true; 3532 } 3533 assert(!Best.hasPendingInjection() && 3534 "All injections should have been done by now!"); 3535 3536 if (Best.TI != PartialIVCondBranch) 3537 PartialIVInfo.InstToDuplicate.clear(); 3538 3539 bool InsertFreeze; 3540 if (auto *SI = dyn_cast<SelectInst>(Best.TI)) { 3541 // If the best candidate is a select, turn it into a branch. Select 3542 // instructions with a poison conditional do not propagate poison, but 3543 // branching on poison causes UB. Insert a freeze on the select 3544 // conditional to prevent UB after turning the select into a branch. 3545 InsertFreeze = !isGuaranteedNotToBeUndefOrPoison( 3546 SI->getCondition(), &AC, L.getLoopPreheader()->getTerminator(), &DT); 3547 Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, &AC); 3548 } else { 3549 // If the best candidate is a guard, turn it into a branch. 3550 if (isGuard(Best.TI)) 3551 Best.TI = 3552 turnGuardIntoBranch(cast<IntrinsicInst>(Best.TI), L, DT, LI, MSSAU); 3553 InsertFreeze = shouldInsertFreeze(L, *Best.TI, DT, AC); 3554 } 3555 3556 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = " << Best.Cost 3557 << ") terminator: " << *Best.TI << "\n"); 3558 unswitchNontrivialInvariants(L, *Best.TI, Best.Invariants, PartialIVInfo, DT, 3559 LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze, 3560 InjectedCondition); 3561 return true; 3562 } 3563 3564 /// Unswitch control flow predicated on loop invariant conditions. 3565 /// 3566 /// This first hoists all branches or switches which are trivial (IE, do not 3567 /// require duplicating any part of the loop) out of the loop body. It then 3568 /// looks at other loop invariant control flows and tries to unswitch those as 3569 /// well by cloning the loop if the result is small enough. 3570 /// 3571 /// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are 3572 /// also updated based on the unswitch. The `MSSA` analysis is also updated if 3573 /// valid (i.e. its use is enabled). 3574 /// 3575 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is 3576 /// true, we will attempt to do non-trivial unswitching as well as trivial 3577 /// unswitching. 3578 /// 3579 /// The `postUnswitch` function will be run after unswitching is complete 3580 /// with information on whether or not the provided loop remains a loop and 3581 /// a list of new sibling loops created. 3582 /// 3583 /// If `SE` is non-null, we will update that analysis based on the unswitching 3584 /// done. 3585 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, 3586 AssumptionCache &AC, AAResults &AA, 3587 TargetTransformInfo &TTI, bool Trivial, 3588 bool NonTrivial, ScalarEvolution *SE, 3589 MemorySSAUpdater *MSSAU, ProfileSummaryInfo *PSI, 3590 BlockFrequencyInfo *BFI, LPMUpdater &LoopUpdater) { 3591 assert(L.isRecursivelyLCSSAForm(DT, LI) && 3592 "Loops must be in LCSSA form before unswitching."); 3593 3594 // Must be in loop simplified form: we need a preheader and dedicated exits. 3595 if (!L.isLoopSimplifyForm()) 3596 return false; 3597 3598 // Try trivial unswitch first before loop over other basic blocks in the loop. 3599 if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) { 3600 // If we unswitched successfully we will want to clean up the loop before 3601 // processing it further so just mark it as unswitched and return. 3602 postUnswitch(L, LoopUpdater, L.getName(), 3603 /*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false, 3604 /*InjectedCondition*/ false, {}); 3605 return true; 3606 } 3607 3608 const Function *F = L.getHeader()->getParent(); 3609 3610 // Check whether we should continue with non-trivial conditions. 3611 // EnableNonTrivialUnswitch: Global variable that forces non-trivial 3612 // unswitching for testing and debugging. 3613 // NonTrivial: Parameter that enables non-trivial unswitching for this 3614 // invocation of the transform. But this should be allowed only 3615 // for targets without branch divergence. 3616 // 3617 // FIXME: If divergence analysis becomes available to a loop 3618 // transform, we should allow unswitching for non-trivial uniform 3619 // branches even on targets that have divergence. 3620 // https://bugs.llvm.org/show_bug.cgi?id=48819 3621 bool ContinueWithNonTrivial = 3622 EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F)); 3623 if (!ContinueWithNonTrivial) 3624 return false; 3625 3626 // Skip non-trivial unswitching for optsize functions. 3627 if (F->hasOptSize()) 3628 return false; 3629 3630 // Returns true if Loop L's loop nest is cold, i.e. if the headers of L, 3631 // of the loops L is nested in, and of the loops nested in L are all cold. 3632 auto IsLoopNestCold = [&](const Loop *L) { 3633 // Check L and all of its parent loops. 3634 auto *Parent = L; 3635 while (Parent) { 3636 if (!PSI->isColdBlock(Parent->getHeader(), BFI)) 3637 return false; 3638 Parent = Parent->getParentLoop(); 3639 } 3640 // Next check all loops nested within L. 3641 SmallVector<const Loop *, 4> Worklist; 3642 Worklist.insert(Worklist.end(), L->getSubLoops().begin(), 3643 L->getSubLoops().end()); 3644 while (!Worklist.empty()) { 3645 auto *CurLoop = Worklist.pop_back_val(); 3646 if (!PSI->isColdBlock(CurLoop->getHeader(), BFI)) 3647 return false; 3648 Worklist.insert(Worklist.end(), CurLoop->getSubLoops().begin(), 3649 CurLoop->getSubLoops().end()); 3650 } 3651 return true; 3652 }; 3653 3654 // Skip cold loops in cold loop nests, as unswitching them brings little 3655 // benefit but increases the code size 3656 if (PSI && PSI->hasProfileSummary() && BFI && IsLoopNestCold(&L)) { 3657 LLVM_DEBUG(dbgs() << " Skip cold loop: " << L << "\n"); 3658 return false; 3659 } 3660 3661 // Perform legality checks. 3662 if (!isSafeForNoNTrivialUnswitching(L, LI)) 3663 return false; 3664 3665 // For non-trivial unswitching, because it often creates new loops, we rely on 3666 // the pass manager to iterate on the loops rather than trying to immediately 3667 // reach a fixed point. There is no substantial advantage to iterating 3668 // internally, and if any of the new loops are simplified enough to contain 3669 // trivial unswitching we want to prefer those. 3670 3671 // Try to unswitch the best invariant condition. We prefer this full unswitch to 3672 // a partial unswitch when possible below the threshold. 3673 if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater)) 3674 return true; 3675 3676 // No other opportunities to unswitch. 3677 return false; 3678 } 3679 3680 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, 3681 LoopStandardAnalysisResults &AR, 3682 LPMUpdater &U) { 3683 Function &F = *L.getHeader()->getParent(); 3684 (void)F; 3685 ProfileSummaryInfo *PSI = nullptr; 3686 if (auto OuterProxy = 3687 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR) 3688 .getCachedResult<ModuleAnalysisManagerFunctionProxy>(F)) 3689 PSI = OuterProxy->getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); 3690 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L 3691 << "\n"); 3692 3693 std::optional<MemorySSAUpdater> MSSAU; 3694 if (AR.MSSA) { 3695 MSSAU = MemorySSAUpdater(AR.MSSA); 3696 if (VerifyMemorySSA) 3697 AR.MSSA->verifyMemorySSA(); 3698 } 3699 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial, 3700 &AR.SE, MSSAU ? &*MSSAU : nullptr, PSI, AR.BFI, U)) 3701 return PreservedAnalyses::all(); 3702 3703 if (AR.MSSA && VerifyMemorySSA) 3704 AR.MSSA->verifyMemorySSA(); 3705 3706 // Historically this pass has had issues with the dominator tree so verify it 3707 // in asserts builds. 3708 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast)); 3709 3710 auto PA = getLoopPassPreservedAnalyses(); 3711 if (AR.MSSA) 3712 PA.preserve<MemorySSAAnalysis>(); 3713 return PA; 3714 } 3715 3716 void SimpleLoopUnswitchPass::printPipeline( 3717 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 3718 static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline( 3719 OS, MapClassName2PassName); 3720 3721 OS << '<'; 3722 OS << (NonTrivial ? "" : "no-") << "nontrivial;"; 3723 OS << (Trivial ? "" : "no-") << "trivial"; 3724 OS << '>'; 3725 } 3726