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