1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 11 #include "llvm/ADT/DenseMap.h" 12 #include "llvm/ADT/STLExtras.h" 13 #include "llvm/ADT/Sequence.h" 14 #include "llvm/ADT/SetVector.h" 15 #include "llvm/ADT/SmallPtrSet.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/ADT/Twine.h" 19 #include "llvm/Analysis/AssumptionCache.h" 20 #include "llvm/Analysis/CFG.h" 21 #include "llvm/Analysis/CodeMetrics.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/Utils/Local.h" 28 #include "llvm/IR/BasicBlock.h" 29 #include "llvm/IR/Constant.h" 30 #include "llvm/IR/Constants.h" 31 #include "llvm/IR/Dominators.h" 32 #include "llvm/IR/Function.h" 33 #include "llvm/IR/InstrTypes.h" 34 #include "llvm/IR/Instruction.h" 35 #include "llvm/IR/Instructions.h" 36 #include "llvm/IR/IntrinsicInst.h" 37 #include "llvm/IR/Use.h" 38 #include "llvm/IR/Value.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/Casting.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/ErrorHandling.h" 43 #include "llvm/Support/GenericDomTree.h" 44 #include "llvm/Support/raw_ostream.h" 45 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 46 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 47 #include "llvm/Transforms/Utils/Cloning.h" 48 #include "llvm/Transforms/Utils/LoopUtils.h" 49 #include "llvm/Transforms/Utils/ValueMapper.h" 50 #include <algorithm> 51 #include <cassert> 52 #include <iterator> 53 #include <numeric> 54 #include <utility> 55 56 #define DEBUG_TYPE "simple-loop-unswitch" 57 58 using namespace llvm; 59 60 STATISTIC(NumBranches, "Number of branches unswitched"); 61 STATISTIC(NumSwitches, "Number of switches unswitched"); 62 STATISTIC(NumTrivial, "Number of unswitches that are trivial"); 63 64 static cl::opt<bool> EnableNonTrivialUnswitch( 65 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden, 66 cl::desc("Forcibly enables non-trivial loop unswitching rather than " 67 "following the configuration passed into the pass.")); 68 69 static cl::opt<int> 70 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, 71 cl::desc("The cost threshold for unswitching a loop.")); 72 73 /// Collect all of the loop invariant input values transitively used by the 74 /// homogeneous instruction graph from a given root. 75 /// 76 /// This essentially walks from a root recursively through loop variant operands 77 /// which have the exact same opcode and finds all inputs which are loop 78 /// invariant. For some operations these can be re-associated and unswitched out 79 /// of the loop entirely. 80 static TinyPtrVector<Value *> 81 collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root, 82 LoopInfo &LI) { 83 assert(!L.isLoopInvariant(&Root) && 84 "Only need to walk the graph if root itself is not invariant."); 85 TinyPtrVector<Value *> Invariants; 86 87 // Build a worklist and recurse through operators collecting invariants. 88 SmallVector<Instruction *, 4> Worklist; 89 SmallPtrSet<Instruction *, 8> Visited; 90 Worklist.push_back(&Root); 91 Visited.insert(&Root); 92 do { 93 Instruction &I = *Worklist.pop_back_val(); 94 for (Value *OpV : I.operand_values()) { 95 // Skip constants as unswitching isn't interesting for them. 96 if (isa<Constant>(OpV)) 97 continue; 98 99 // Add it to our result if loop invariant. 100 if (L.isLoopInvariant(OpV)) { 101 Invariants.push_back(OpV); 102 continue; 103 } 104 105 // If not an instruction with the same opcode, nothing we can do. 106 Instruction *OpI = dyn_cast<Instruction>(OpV); 107 if (!OpI || OpI->getOpcode() != Root.getOpcode()) 108 continue; 109 110 // Visit this operand. 111 if (Visited.insert(OpI).second) 112 Worklist.push_back(OpI); 113 } 114 } while (!Worklist.empty()); 115 116 return Invariants; 117 } 118 119 static void replaceLoopInvariantUses(Loop &L, Value *Invariant, 120 Constant &Replacement) { 121 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?"); 122 123 // Replace uses of LIC in the loop with the given constant. 124 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); UI != UE;) { 125 // Grab the use and walk past it so we can clobber it in the use list. 126 Use *U = &*UI++; 127 Instruction *UserI = dyn_cast<Instruction>(U->getUser()); 128 129 // Replace this use within the loop body. 130 if (UserI && L.contains(UserI)) 131 U->set(&Replacement); 132 } 133 } 134 135 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial 136 /// incoming values along this edge. 137 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB, 138 BasicBlock &ExitBB) { 139 for (Instruction &I : ExitBB) { 140 auto *PN = dyn_cast<PHINode>(&I); 141 if (!PN) 142 // No more PHIs to check. 143 return true; 144 145 // If the incoming value for this edge isn't loop invariant the unswitch 146 // won't be trivial. 147 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))) 148 return false; 149 } 150 llvm_unreachable("Basic blocks should never be empty!"); 151 } 152 153 /// Insert code to test a set of loop invariant values, and conditionally branch 154 /// on them. 155 static void buildPartialUnswitchConditionalBranch(BasicBlock &BB, 156 ArrayRef<Value *> Invariants, 157 bool Direction, 158 BasicBlock &UnswitchedSucc, 159 BasicBlock &NormalSucc) { 160 IRBuilder<> IRB(&BB); 161 Value *Cond = Invariants.front(); 162 for (Value *Invariant : 163 make_range(std::next(Invariants.begin()), Invariants.end())) 164 if (Direction) 165 Cond = IRB.CreateOr(Cond, Invariant); 166 else 167 Cond = IRB.CreateAnd(Cond, Invariant); 168 169 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, 170 Direction ? &NormalSucc : &UnswitchedSucc); 171 } 172 173 /// Rewrite the PHI nodes in an unswitched loop exit basic block. 174 /// 175 /// Requires that the loop exit and unswitched basic block are the same, and 176 /// that the exiting block was a unique predecessor of that block. Rewrites the 177 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial 178 /// PHI nodes from the old preheader that now contains the unswitched 179 /// terminator. 180 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, 181 BasicBlock &OldExitingBB, 182 BasicBlock &OldPH) { 183 for (PHINode &PN : UnswitchedBB.phis()) { 184 // When the loop exit is directly unswitched we just need to update the 185 // incoming basic block. We loop to handle weird cases with repeated 186 // incoming blocks, but expect to typically only have one operand here. 187 for (auto i : seq<int>(0, PN.getNumOperands())) { 188 assert(PN.getIncomingBlock(i) == &OldExitingBB && 189 "Found incoming block different from unique predecessor!"); 190 PN.setIncomingBlock(i, &OldPH); 191 } 192 } 193 } 194 195 /// Rewrite the PHI nodes in the loop exit basic block and the split off 196 /// unswitched block. 197 /// 198 /// Because the exit block remains an exit from the loop, this rewrites the 199 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI 200 /// nodes into the unswitched basic block to select between the value in the 201 /// old preheader and the loop exit. 202 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, 203 BasicBlock &UnswitchedBB, 204 BasicBlock &OldExitingBB, 205 BasicBlock &OldPH, 206 bool FullUnswitch) { 207 assert(&ExitBB != &UnswitchedBB && 208 "Must have different loop exit and unswitched blocks!"); 209 Instruction *InsertPt = &*UnswitchedBB.begin(); 210 for (PHINode &PN : ExitBB.phis()) { 211 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2, 212 PN.getName() + ".split", InsertPt); 213 214 // Walk backwards over the old PHI node's inputs to minimize the cost of 215 // removing each one. We have to do this weird loop manually so that we 216 // create the same number of new incoming edges in the new PHI as we expect 217 // each case-based edge to be included in the unswitched switch in some 218 // cases. 219 // FIXME: This is really, really gross. It would be much cleaner if LLVM 220 // allowed us to create a single entry for a predecessor block without 221 // having separate entries for each "edge" even though these edges are 222 // required to produce identical results. 223 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) { 224 if (PN.getIncomingBlock(i) != &OldExitingBB) 225 continue; 226 227 Value *Incoming = PN.getIncomingValue(i); 228 if (FullUnswitch) 229 // No more edge from the old exiting block to the exit block. 230 PN.removeIncomingValue(i); 231 232 NewPN->addIncoming(Incoming, &OldPH); 233 } 234 235 // Now replace the old PHI with the new one and wire the old one in as an 236 // input to the new one. 237 PN.replaceAllUsesWith(NewPN); 238 NewPN->addIncoming(&PN, &ExitBB); 239 } 240 } 241 242 /// Hoist the current loop up to the innermost loop containing a remaining exit. 243 /// 244 /// Because we've removed an exit from the loop, we may have changed the set of 245 /// loops reachable and need to move the current loop up the loop nest or even 246 /// to an entirely separate nest. 247 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader, 248 DominatorTree &DT, LoopInfo &LI) { 249 // If the loop is already at the top level, we can't hoist it anywhere. 250 Loop *OldParentL = L.getParentLoop(); 251 if (!OldParentL) 252 return; 253 254 SmallVector<BasicBlock *, 4> Exits; 255 L.getExitBlocks(Exits); 256 Loop *NewParentL = nullptr; 257 for (auto *ExitBB : Exits) 258 if (Loop *ExitL = LI.getLoopFor(ExitBB)) 259 if (!NewParentL || NewParentL->contains(ExitL)) 260 NewParentL = ExitL; 261 262 if (NewParentL == OldParentL) 263 return; 264 265 // The new parent loop (if different) should always contain the old one. 266 if (NewParentL) 267 assert(NewParentL->contains(OldParentL) && 268 "Can only hoist this loop up the nest!"); 269 270 // The preheader will need to move with the body of this loop. However, 271 // because it isn't in this loop we also need to update the primary loop map. 272 assert(OldParentL == LI.getLoopFor(&Preheader) && 273 "Parent loop of this loop should contain this loop's preheader!"); 274 LI.changeLoopFor(&Preheader, NewParentL); 275 276 // Remove this loop from its old parent. 277 OldParentL->removeChildLoop(&L); 278 279 // Add the loop either to the new parent or as a top-level loop. 280 if (NewParentL) 281 NewParentL->addChildLoop(&L); 282 else 283 LI.addTopLevelLoop(&L); 284 285 // Remove this loops blocks from the old parent and every other loop up the 286 // nest until reaching the new parent. Also update all of these 287 // no-longer-containing loops to reflect the nesting change. 288 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL; 289 OldContainingL = OldContainingL->getParentLoop()) { 290 llvm::erase_if(OldContainingL->getBlocksVector(), 291 [&](const BasicBlock *BB) { 292 return BB == &Preheader || L.contains(BB); 293 }); 294 295 OldContainingL->getBlocksSet().erase(&Preheader); 296 for (BasicBlock *BB : L.blocks()) 297 OldContainingL->getBlocksSet().erase(BB); 298 299 // Because we just hoisted a loop out of this one, we have essentially 300 // created new exit paths from it. That means we need to form LCSSA PHI 301 // nodes for values used in the no-longer-nested loop. 302 formLCSSA(*OldContainingL, DT, &LI, nullptr); 303 304 // We shouldn't need to form dedicated exits because the exit introduced 305 // here is the (just split by unswitching) preheader. However, after trivial 306 // unswitching it is possible to get new non-dedicated exits out of parent 307 // loop so let's conservatively form dedicated exit blocks and figure out 308 // if we can optimize later. 309 formDedicatedExitBlocks(OldContainingL, &DT, &LI, /*PreserveLCSSA*/ true); 310 } 311 } 312 313 /// Unswitch a trivial branch if the condition is loop invariant. 314 /// 315 /// This routine should only be called when loop code leading to the branch has 316 /// been validated as trivial (no side effects). This routine checks if the 317 /// condition is invariant and one of the successors is a loop exit. This 318 /// allows us to unswitch without duplicating the loop, making it trivial. 319 /// 320 /// If this routine fails to unswitch the branch it returns false. 321 /// 322 /// If the branch can be unswitched, this routine splits the preheader and 323 /// hoists the branch above that split. Preserves loop simplified form 324 /// (splitting the exit block as necessary). It simplifies the branch within 325 /// the loop to an unconditional branch but doesn't remove it entirely. Further 326 /// cleanup can be done with some simplify-cfg like pass. 327 /// 328 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 329 /// invalidated by this. 330 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, 331 LoopInfo &LI, ScalarEvolution *SE) { 332 assert(BI.isConditional() && "Can only unswitch a conditional branch!"); 333 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n"); 334 335 // The loop invariant values that we want to unswitch. 336 TinyPtrVector<Value *> Invariants; 337 338 // When true, we're fully unswitching the branch rather than just unswitching 339 // some input conditions to the branch. 340 bool FullUnswitch = false; 341 342 if (L.isLoopInvariant(BI.getCondition())) { 343 Invariants.push_back(BI.getCondition()); 344 FullUnswitch = true; 345 } else { 346 if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition())) 347 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI); 348 if (Invariants.empty()) 349 // Couldn't find invariant inputs! 350 return false; 351 } 352 353 // Check that one of the branch's successors exits, and which one. 354 bool ExitDirection = true; 355 int LoopExitSuccIdx = 0; 356 auto *LoopExitBB = BI.getSuccessor(0); 357 if (L.contains(LoopExitBB)) { 358 ExitDirection = false; 359 LoopExitSuccIdx = 1; 360 LoopExitBB = BI.getSuccessor(1); 361 if (L.contains(LoopExitBB)) 362 return false; 363 } 364 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); 365 auto *ParentBB = BI.getParent(); 366 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) 367 return false; 368 369 // When unswitching only part of the branch's condition, we need the exit 370 // block to be reached directly from the partially unswitched input. This can 371 // be done when the exit block is along the true edge and the branch condition 372 // is a graph of `or` operations, or the exit block is along the false edge 373 // and the condition is a graph of `and` operations. 374 if (!FullUnswitch) { 375 if (ExitDirection) { 376 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::Or) 377 return false; 378 } else { 379 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::And) 380 return false; 381 } 382 } 383 384 LLVM_DEBUG({ 385 dbgs() << " unswitching trivial invariant conditions for: " << BI 386 << "\n"; 387 for (Value *Invariant : Invariants) { 388 dbgs() << " " << *Invariant << " == true"; 389 if (Invariant != Invariants.back()) 390 dbgs() << " ||"; 391 dbgs() << "\n"; 392 } 393 }); 394 395 // If we have scalar evolutions, we need to invalidate them including this 396 // loop and the loop containing the exit block. 397 if (SE) { 398 if (Loop *ExitL = LI.getLoopFor(LoopExitBB)) 399 SE->forgetLoop(ExitL); 400 else 401 // Forget the entire nest as this exits the entire nest. 402 SE->forgetTopmostLoop(&L); 403 } 404 405 // Split the preheader, so that we know that there is a safe place to insert 406 // the conditional branch. We will change the preheader to have a conditional 407 // branch on LoopCond. 408 BasicBlock *OldPH = L.getLoopPreheader(); 409 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI); 410 411 // Now that we have a place to insert the conditional branch, create a place 412 // to branch to: this is the exit block out of the loop that we are 413 // unswitching. We need to split this if there are other loop predecessors. 414 // Because the loop is in simplified form, *any* other predecessor is enough. 415 BasicBlock *UnswitchedBB; 416 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) { 417 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() && 418 "A branch's parent isn't a predecessor!"); 419 UnswitchedBB = LoopExitBB; 420 } else { 421 UnswitchedBB = SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI); 422 } 423 424 // Actually move the invariant uses into the unswitched position. If possible, 425 // we do this by moving the instructions, but when doing partial unswitching 426 // we do it by building a new merge of the values in the unswitched position. 427 OldPH->getTerminator()->eraseFromParent(); 428 if (FullUnswitch) { 429 // If fully unswitching, we can use the existing branch instruction. 430 // Splice it into the old PH to gate reaching the new preheader and re-point 431 // its successors. 432 OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(), 433 BI); 434 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); 435 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); 436 437 // Create a new unconditional branch that will continue the loop as a new 438 // terminator. 439 BranchInst::Create(ContinueBB, ParentBB); 440 } else { 441 // Only unswitching a subset of inputs to the condition, so we will need to 442 // build a new branch that merges the invariant inputs. 443 if (ExitDirection) 444 assert(cast<Instruction>(BI.getCondition())->getOpcode() == 445 Instruction::Or && 446 "Must have an `or` of `i1`s for the condition!"); 447 else 448 assert(cast<Instruction>(BI.getCondition())->getOpcode() == 449 Instruction::And && 450 "Must have an `and` of `i1`s for the condition!"); 451 buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection, 452 *UnswitchedBB, *NewPH); 453 } 454 455 // Rewrite the relevant PHI nodes. 456 if (UnswitchedBB == LoopExitBB) 457 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); 458 else 459 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, 460 *ParentBB, *OldPH, FullUnswitch); 461 462 // Now we need to update the dominator tree. 463 SmallVector<DominatorTree::UpdateType, 2> DTUpdates; 464 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedBB}); 465 if (FullUnswitch) 466 DTUpdates.push_back({DT.Delete, ParentBB, LoopExitBB}); 467 DT.applyUpdates(DTUpdates); 468 469 // The constant we can replace all of our invariants with inside the loop 470 // body. If any of the invariants have a value other than this the loop won't 471 // be entered. 472 ConstantInt *Replacement = ExitDirection 473 ? ConstantInt::getFalse(BI.getContext()) 474 : ConstantInt::getTrue(BI.getContext()); 475 476 // Since this is an i1 condition we can also trivially replace uses of it 477 // within the loop with a constant. 478 for (Value *Invariant : Invariants) 479 replaceLoopInvariantUses(L, Invariant, *Replacement); 480 481 // If this was full unswitching, we may have changed the nesting relationship 482 // for this loop so hoist it to its correct parent if needed. 483 if (FullUnswitch) 484 hoistLoopToNewParent(L, *NewPH, DT, LI); 485 486 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n"); 487 ++NumTrivial; 488 ++NumBranches; 489 return true; 490 } 491 492 /// Unswitch a trivial switch if the condition is loop invariant. 493 /// 494 /// This routine should only be called when loop code leading to the switch has 495 /// been validated as trivial (no side effects). This routine checks if the 496 /// condition is invariant and that at least one of the successors is a loop 497 /// exit. This allows us to unswitch without duplicating the loop, making it 498 /// trivial. 499 /// 500 /// If this routine fails to unswitch the switch it returns false. 501 /// 502 /// If the switch can be unswitched, this routine splits the preheader and 503 /// copies the switch above that split. If the default case is one of the 504 /// exiting cases, it copies the non-exiting cases and points them at the new 505 /// preheader. If the default case is not exiting, it copies the exiting cases 506 /// and points the default at the preheader. It preserves loop simplified form 507 /// (splitting the exit blocks as necessary). It simplifies the switch within 508 /// the loop by removing now-dead cases. If the default case is one of those 509 /// unswitched, it replaces its destination with a new basic block containing 510 /// only unreachable. Such basic blocks, while technically loop exits, are not 511 /// considered for unswitching so this is a stable transform and the same 512 /// switch will not be revisited. If after unswitching there is only a single 513 /// in-loop successor, the switch is further simplified to an unconditional 514 /// branch. Still more cleanup can be done with some simplify-cfg like pass. 515 /// 516 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 517 /// invalidated by this. 518 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, 519 LoopInfo &LI, ScalarEvolution *SE) { 520 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n"); 521 Value *LoopCond = SI.getCondition(); 522 523 // If this isn't switching on an invariant condition, we can't unswitch it. 524 if (!L.isLoopInvariant(LoopCond)) 525 return false; 526 527 auto *ParentBB = SI.getParent(); 528 529 SmallVector<int, 4> ExitCaseIndices; 530 for (auto Case : SI.cases()) { 531 auto *SuccBB = Case.getCaseSuccessor(); 532 if (!L.contains(SuccBB) && 533 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB)) 534 ExitCaseIndices.push_back(Case.getCaseIndex()); 535 } 536 BasicBlock *DefaultExitBB = nullptr; 537 if (!L.contains(SI.getDefaultDest()) && 538 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) && 539 !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator())) 540 DefaultExitBB = SI.getDefaultDest(); 541 else if (ExitCaseIndices.empty()) 542 return false; 543 544 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n"); 545 546 // We may need to invalidate SCEVs for the outermost loop reached by any of 547 // the exits. 548 Loop *OuterL = &L; 549 550 if (DefaultExitBB) { 551 // Clear out the default destination temporarily to allow accurate 552 // predecessor lists to be examined below. 553 SI.setDefaultDest(nullptr); 554 // Check the loop containing this exit. 555 Loop *ExitL = LI.getLoopFor(DefaultExitBB); 556 if (!ExitL || ExitL->contains(OuterL)) 557 OuterL = ExitL; 558 } 559 560 // Store the exit cases into a separate data structure and remove them from 561 // the switch. 562 SmallVector<std::pair<ConstantInt *, BasicBlock *>, 4> ExitCases; 563 ExitCases.reserve(ExitCaseIndices.size()); 564 // We walk the case indices backwards so that we remove the last case first 565 // and don't disrupt the earlier indices. 566 for (unsigned Index : reverse(ExitCaseIndices)) { 567 auto CaseI = SI.case_begin() + Index; 568 // Compute the outer loop from this exit. 569 Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor()); 570 if (!ExitL || ExitL->contains(OuterL)) 571 OuterL = ExitL; 572 // Save the value of this case. 573 ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()}); 574 // Delete the unswitched cases. 575 SI.removeCase(CaseI); 576 } 577 578 if (SE) { 579 if (OuterL) 580 SE->forgetLoop(OuterL); 581 else 582 SE->forgetTopmostLoop(&L); 583 } 584 585 // Check if after this all of the remaining cases point at the same 586 // successor. 587 BasicBlock *CommonSuccBB = nullptr; 588 if (SI.getNumCases() > 0 && 589 std::all_of(std::next(SI.case_begin()), SI.case_end(), 590 [&SI](const SwitchInst::CaseHandle &Case) { 591 return Case.getCaseSuccessor() == 592 SI.case_begin()->getCaseSuccessor(); 593 })) 594 CommonSuccBB = SI.case_begin()->getCaseSuccessor(); 595 if (!DefaultExitBB) { 596 // If we're not unswitching the default, we need it to match any cases to 597 // have a common successor or if we have no cases it is the common 598 // successor. 599 if (SI.getNumCases() == 0) 600 CommonSuccBB = SI.getDefaultDest(); 601 else if (SI.getDefaultDest() != CommonSuccBB) 602 CommonSuccBB = nullptr; 603 } 604 605 // Split the preheader, so that we know that there is a safe place to insert 606 // the switch. 607 BasicBlock *OldPH = L.getLoopPreheader(); 608 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI); 609 OldPH->getTerminator()->eraseFromParent(); 610 611 // Now add the unswitched switch. 612 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); 613 614 // Rewrite the IR for the unswitched basic blocks. This requires two steps. 615 // First, we split any exit blocks with remaining in-loop predecessors. Then 616 // we update the PHIs in one of two ways depending on if there was a split. 617 // We walk in reverse so that we split in the same order as the cases 618 // appeared. This is purely for convenience of reading the resulting IR, but 619 // it doesn't cost anything really. 620 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; 621 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; 622 // Handle the default exit if necessary. 623 // FIXME: It'd be great if we could merge this with the loop below but LLVM's 624 // ranges aren't quite powerful enough yet. 625 if (DefaultExitBB) { 626 if (pred_empty(DefaultExitBB)) { 627 UnswitchedExitBBs.insert(DefaultExitBB); 628 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); 629 } else { 630 auto *SplitBB = 631 SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI); 632 rewritePHINodesForExitAndUnswitchedBlocks( 633 *DefaultExitBB, *SplitBB, *ParentBB, *OldPH, /*FullUnswitch*/ true); 634 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; 635 } 636 } 637 // Note that we must use a reference in the for loop so that we update the 638 // container. 639 for (auto &CasePair : reverse(ExitCases)) { 640 // Grab a reference to the exit block in the pair so that we can update it. 641 BasicBlock *ExitBB = CasePair.second; 642 643 // If this case is the last edge into the exit block, we can simply reuse it 644 // as it will no longer be a loop exit. No mapping necessary. 645 if (pred_empty(ExitBB)) { 646 // Only rewrite once. 647 if (UnswitchedExitBBs.insert(ExitBB).second) 648 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); 649 continue; 650 } 651 652 // Otherwise we need to split the exit block so that we retain an exit 653 // block from the loop and a target for the unswitched condition. 654 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; 655 if (!SplitExitBB) { 656 // If this is the first time we see this, do the split and remember it. 657 SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI); 658 rewritePHINodesForExitAndUnswitchedBlocks( 659 *ExitBB, *SplitExitBB, *ParentBB, *OldPH, /*FullUnswitch*/ true); 660 } 661 // Update the case pair to point to the split block. 662 CasePair.second = SplitExitBB; 663 } 664 665 // Now add the unswitched cases. We do this in reverse order as we built them 666 // in reverse order. 667 for (auto CasePair : reverse(ExitCases)) { 668 ConstantInt *CaseVal = CasePair.first; 669 BasicBlock *UnswitchedBB = CasePair.second; 670 671 NewSI->addCase(CaseVal, UnswitchedBB); 672 } 673 674 // If the default was unswitched, re-point it and add explicit cases for 675 // entering the loop. 676 if (DefaultExitBB) { 677 NewSI->setDefaultDest(DefaultExitBB); 678 679 // We removed all the exit cases, so we just copy the cases to the 680 // unswitched switch. 681 for (auto Case : SI.cases()) 682 NewSI->addCase(Case.getCaseValue(), NewPH); 683 } 684 685 // If we ended up with a common successor for every path through the switch 686 // after unswitching, rewrite it to an unconditional branch to make it easy 687 // to recognize. Otherwise we potentially have to recognize the default case 688 // pointing at unreachable and other complexity. 689 if (CommonSuccBB) { 690 BasicBlock *BB = SI.getParent(); 691 // We may have had multiple edges to this common successor block, so remove 692 // them as predecessors. We skip the first one, either the default or the 693 // actual first case. 694 bool SkippedFirst = DefaultExitBB == nullptr; 695 for (auto Case : SI.cases()) { 696 assert(Case.getCaseSuccessor() == CommonSuccBB && 697 "Non-common successor!"); 698 (void)Case; 699 if (!SkippedFirst) { 700 SkippedFirst = true; 701 continue; 702 } 703 CommonSuccBB->removePredecessor(BB, 704 /*DontDeleteUselessPHIs*/ true); 705 } 706 // Now nuke the switch and replace it with a direct branch. 707 SI.eraseFromParent(); 708 BranchInst::Create(CommonSuccBB, BB); 709 } else if (DefaultExitBB) { 710 assert(SI.getNumCases() > 0 && 711 "If we had no cases we'd have a common successor!"); 712 // Move the last case to the default successor. This is valid as if the 713 // default got unswitched it cannot be reached. This has the advantage of 714 // being simple and keeping the number of edges from this switch to 715 // successors the same, and avoiding any PHI update complexity. 716 auto LastCaseI = std::prev(SI.case_end()); 717 SI.setDefaultDest(LastCaseI->getCaseSuccessor()); 718 SI.removeCase(LastCaseI); 719 } 720 721 // Walk the unswitched exit blocks and the unswitched split blocks and update 722 // the dominator tree based on the CFG edits. While we are walking unordered 723 // containers here, the API for applyUpdates takes an unordered list of 724 // updates and requires them to not contain duplicates. 725 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 726 for (auto *UnswitchedExitBB : UnswitchedExitBBs) { 727 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB}); 728 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB}); 729 } 730 for (auto SplitUnswitchedPair : SplitExitBBMap) { 731 auto *UnswitchedBB = SplitUnswitchedPair.second; 732 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedBB}); 733 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedBB}); 734 } 735 DT.applyUpdates(DTUpdates); 736 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 737 738 // We may have changed the nesting relationship for this loop so hoist it to 739 // its correct parent if needed. 740 hoistLoopToNewParent(L, *NewPH, DT, LI); 741 742 ++NumTrivial; 743 ++NumSwitches; 744 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n"); 745 return true; 746 } 747 748 /// This routine scans the loop to find a branch or switch which occurs before 749 /// any side effects occur. These can potentially be unswitched without 750 /// duplicating the loop. If a branch or switch is successfully unswitched the 751 /// scanning continues to see if subsequent branches or switches have become 752 /// trivial. Once all trivial candidates have been unswitched, this routine 753 /// returns. 754 /// 755 /// The return value indicates whether anything was unswitched (and therefore 756 /// changed). 757 /// 758 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 759 /// invalidated by this. 760 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, 761 LoopInfo &LI, ScalarEvolution *SE) { 762 bool Changed = false; 763 764 // If loop header has only one reachable successor we should keep looking for 765 // trivial condition candidates in the successor as well. An alternative is 766 // to constant fold conditions and merge successors into loop header (then we 767 // only need to check header's terminator). The reason for not doing this in 768 // LoopUnswitch pass is that it could potentially break LoopPassManager's 769 // invariants. Folding dead branches could either eliminate the current loop 770 // or make other loops unreachable. LCSSA form might also not be preserved 771 // after deleting branches. The following code keeps traversing loop header's 772 // successors until it finds the trivial condition candidate (condition that 773 // is not a constant). Since unswitching generates branches with constant 774 // conditions, this scenario could be very common in practice. 775 BasicBlock *CurrentBB = L.getHeader(); 776 SmallPtrSet<BasicBlock *, 8> Visited; 777 Visited.insert(CurrentBB); 778 do { 779 // Check if there are any side-effecting instructions (e.g. stores, calls, 780 // volatile loads) in the part of the loop that the code *would* execute 781 // without unswitching. 782 if (llvm::any_of(*CurrentBB, 783 [](Instruction &I) { return I.mayHaveSideEffects(); })) 784 return Changed; 785 786 TerminatorInst *CurrentTerm = CurrentBB->getTerminator(); 787 788 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) { 789 // Don't bother trying to unswitch past a switch with a constant 790 // condition. This should be removed prior to running this pass by 791 // simplify-cfg. 792 if (isa<Constant>(SI->getCondition())) 793 return Changed; 794 795 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE)) 796 // Couldn't unswitch this one so we're done. 797 return Changed; 798 799 // Mark that we managed to unswitch something. 800 Changed = true; 801 802 // If unswitching turned the terminator into an unconditional branch then 803 // we can continue. The unswitching logic specifically works to fold any 804 // cases it can into an unconditional branch to make it easier to 805 // recognize here. 806 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); 807 if (!BI || BI->isConditional()) 808 return Changed; 809 810 CurrentBB = BI->getSuccessor(0); 811 continue; 812 } 813 814 auto *BI = dyn_cast<BranchInst>(CurrentTerm); 815 if (!BI) 816 // We do not understand other terminator instructions. 817 return Changed; 818 819 // Don't bother trying to unswitch past an unconditional branch or a branch 820 // with a constant value. These should be removed by simplify-cfg prior to 821 // running this pass. 822 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 823 return Changed; 824 825 // Found a trivial condition candidate: non-foldable conditional branch. If 826 // we fail to unswitch this, we can't do anything else that is trivial. 827 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE)) 828 return Changed; 829 830 // Mark that we managed to unswitch something. 831 Changed = true; 832 833 // If we only unswitched some of the conditions feeding the branch, we won't 834 // have collapsed it to a single successor. 835 BI = cast<BranchInst>(CurrentBB->getTerminator()); 836 if (BI->isConditional()) 837 return Changed; 838 839 // Follow the newly unconditional branch into its successor. 840 CurrentBB = BI->getSuccessor(0); 841 842 // When continuing, if we exit the loop or reach a previous visited block, 843 // then we can not reach any trivial condition candidates (unfoldable 844 // branch instructions or switch instructions) and no unswitch can happen. 845 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second); 846 847 return Changed; 848 } 849 850 /// Build the cloned blocks for an unswitched copy of the given loop. 851 /// 852 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and 853 /// after the split block (`SplitBB`) that will be used to select between the 854 /// cloned and original loop. 855 /// 856 /// This routine handles cloning all of the necessary loop blocks and exit 857 /// blocks including rewriting their instructions and the relevant PHI nodes. 858 /// Any loop blocks or exit blocks which are dominated by a different successor 859 /// than the one for this clone of the loop blocks can be trivially skipped. We 860 /// use the `DominatingSucc` map to determine whether a block satisfies that 861 /// property with a simple map lookup. 862 /// 863 /// It also correctly creates the unconditional branch in the cloned 864 /// unswitched parent block to only point at the unswitched successor. 865 /// 866 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit 867 /// block splitting is correctly reflected in `LoopInfo`, essentially all of 868 /// the cloned blocks (and their loops) are left without full `LoopInfo` 869 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned 870 /// blocks to them but doesn't create the cloned `DominatorTree` structure and 871 /// instead the caller must recompute an accurate DT. It *does* correctly 872 /// update the `AssumptionCache` provided in `AC`. 873 static BasicBlock *buildClonedLoopBlocks( 874 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, 875 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB, 876 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, 877 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc, 878 ValueToValueMapTy &VMap, 879 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC, 880 DominatorTree &DT, LoopInfo &LI) { 881 SmallVector<BasicBlock *, 4> NewBlocks; 882 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size()); 883 884 // We will need to clone a bunch of blocks, wrap up the clone operation in 885 // a helper. 886 auto CloneBlock = [&](BasicBlock *OldBB) { 887 // Clone the basic block and insert it before the new preheader. 888 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent()); 889 NewBB->moveBefore(LoopPH); 890 891 // Record this block and the mapping. 892 NewBlocks.push_back(NewBB); 893 VMap[OldBB] = NewBB; 894 895 return NewBB; 896 }; 897 898 // We skip cloning blocks when they have a dominating succ that is not the 899 // succ we are cloning for. 900 auto SkipBlock = [&](BasicBlock *BB) { 901 auto It = DominatingSucc.find(BB); 902 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB; 903 }; 904 905 // First, clone the preheader. 906 auto *ClonedPH = CloneBlock(LoopPH); 907 908 // Then clone all the loop blocks, skipping the ones that aren't necessary. 909 for (auto *LoopBB : L.blocks()) 910 if (!SkipBlock(LoopBB)) 911 CloneBlock(LoopBB); 912 913 // Split all the loop exit edges so that when we clone the exit blocks, if 914 // any of the exit blocks are *also* a preheader for some other loop, we 915 // don't create multiple predecessors entering the loop header. 916 for (auto *ExitBB : ExitBlocks) { 917 if (SkipBlock(ExitBB)) 918 continue; 919 920 // When we are going to clone an exit, we don't need to clone all the 921 // instructions in the exit block and we want to ensure we have an easy 922 // place to merge the CFG, so split the exit first. This is always safe to 923 // do because there cannot be any non-loop predecessors of a loop exit in 924 // loop simplified form. 925 auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI); 926 927 // Rearrange the names to make it easier to write test cases by having the 928 // exit block carry the suffix rather than the merge block carrying the 929 // suffix. 930 MergeBB->takeName(ExitBB); 931 ExitBB->setName(Twine(MergeBB->getName()) + ".split"); 932 933 // Now clone the original exit block. 934 auto *ClonedExitBB = CloneBlock(ExitBB); 935 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && 936 "Exit block should have been split to have one successor!"); 937 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && 938 "Cloned exit block has the wrong successor!"); 939 940 // Remap any cloned instructions and create a merge phi node for them. 941 for (auto ZippedInsts : llvm::zip_first( 942 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())), 943 llvm::make_range(ClonedExitBB->begin(), 944 std::prev(ClonedExitBB->end())))) { 945 Instruction &I = std::get<0>(ZippedInsts); 946 Instruction &ClonedI = std::get<1>(ZippedInsts); 947 948 // The only instructions in the exit block should be PHI nodes and 949 // potentially a landing pad. 950 assert( 951 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && 952 "Bad instruction in exit block!"); 953 // We should have a value map between the instruction and its clone. 954 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!"); 955 956 auto *MergePN = 957 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi", 958 &*MergeBB->getFirstInsertionPt()); 959 I.replaceAllUsesWith(MergePN); 960 MergePN->addIncoming(&I, ExitBB); 961 MergePN->addIncoming(&ClonedI, ClonedExitBB); 962 } 963 } 964 965 // Rewrite the instructions in the cloned blocks to refer to the instructions 966 // in the cloned blocks. We have to do this as a second pass so that we have 967 // everything available. Also, we have inserted new instructions which may 968 // include assume intrinsics, so we update the assumption cache while 969 // processing this. 970 for (auto *ClonedBB : NewBlocks) 971 for (Instruction &I : *ClonedBB) { 972 RemapInstruction(&I, VMap, 973 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 974 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 975 if (II->getIntrinsicID() == Intrinsic::assume) 976 AC.registerAssumption(II); 977 } 978 979 // Update any PHI nodes in the cloned successors of the skipped blocks to not 980 // have spurious incoming values. 981 for (auto *LoopBB : L.blocks()) 982 if (SkipBlock(LoopBB)) 983 for (auto *SuccBB : successors(LoopBB)) 984 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB))) 985 for (PHINode &PN : ClonedSuccBB->phis()) 986 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false); 987 988 // Remove the cloned parent as a predecessor of any successor we ended up 989 // cloning other than the unswitched one. 990 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB)); 991 for (auto *SuccBB : successors(ParentBB)) { 992 if (SuccBB == UnswitchedSuccBB) 993 continue; 994 995 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)); 996 if (!ClonedSuccBB) 997 continue; 998 999 ClonedSuccBB->removePredecessor(ClonedParentBB, 1000 /*DontDeleteUselessPHIs*/ true); 1001 } 1002 1003 // Replace the cloned branch with an unconditional branch to the cloned 1004 // unswitched successor. 1005 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB)); 1006 ClonedParentBB->getTerminator()->eraseFromParent(); 1007 BranchInst::Create(ClonedSuccBB, ClonedParentBB); 1008 1009 // If there are duplicate entries in the PHI nodes because of multiple edges 1010 // to the unswitched successor, we need to nuke all but one as we replaced it 1011 // with a direct branch. 1012 for (PHINode &PN : ClonedSuccBB->phis()) { 1013 bool Found = false; 1014 // Loop over the incoming operands backwards so we can easily delete as we 1015 // go without invalidating the index. 1016 for (int i = PN.getNumOperands() - 1; i >= 0; --i) { 1017 if (PN.getIncomingBlock(i) != ClonedParentBB) 1018 continue; 1019 if (!Found) { 1020 Found = true; 1021 continue; 1022 } 1023 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false); 1024 } 1025 } 1026 1027 // Record the domtree updates for the new blocks. 1028 SmallPtrSet<BasicBlock *, 4> SuccSet; 1029 for (auto *ClonedBB : NewBlocks) { 1030 for (auto *SuccBB : successors(ClonedBB)) 1031 if (SuccSet.insert(SuccBB).second) 1032 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB}); 1033 SuccSet.clear(); 1034 } 1035 1036 return ClonedPH; 1037 } 1038 1039 /// Recursively clone the specified loop and all of its children. 1040 /// 1041 /// The target parent loop for the clone should be provided, or can be null if 1042 /// the clone is a top-level loop. While cloning, all the blocks are mapped 1043 /// with the provided value map. The entire original loop must be present in 1044 /// the value map. The cloned loop is returned. 1045 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, 1046 const ValueToValueMapTy &VMap, LoopInfo &LI) { 1047 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) { 1048 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!"); 1049 ClonedL.reserveBlocks(OrigL.getNumBlocks()); 1050 for (auto *BB : OrigL.blocks()) { 1051 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB)); 1052 ClonedL.addBlockEntry(ClonedBB); 1053 if (LI.getLoopFor(BB) == &OrigL) 1054 LI.changeLoopFor(ClonedBB, &ClonedL); 1055 } 1056 }; 1057 1058 // We specially handle the first loop because it may get cloned into 1059 // a different parent and because we most commonly are cloning leaf loops. 1060 Loop *ClonedRootL = LI.AllocateLoop(); 1061 if (RootParentL) 1062 RootParentL->addChildLoop(ClonedRootL); 1063 else 1064 LI.addTopLevelLoop(ClonedRootL); 1065 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL); 1066 1067 if (OrigRootL.empty()) 1068 return ClonedRootL; 1069 1070 // If we have a nest, we can quickly clone the entire loop nest using an 1071 // iterative approach because it is a tree. We keep the cloned parent in the 1072 // data structure to avoid repeatedly querying through a map to find it. 1073 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone; 1074 // Build up the loops to clone in reverse order as we'll clone them from the 1075 // back. 1076 for (Loop *ChildL : llvm::reverse(OrigRootL)) 1077 LoopsToClone.push_back({ClonedRootL, ChildL}); 1078 do { 1079 Loop *ClonedParentL, *L; 1080 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val(); 1081 Loop *ClonedL = LI.AllocateLoop(); 1082 ClonedParentL->addChildLoop(ClonedL); 1083 AddClonedBlocksToLoop(*L, *ClonedL); 1084 for (Loop *ChildL : llvm::reverse(*L)) 1085 LoopsToClone.push_back({ClonedL, ChildL}); 1086 } while (!LoopsToClone.empty()); 1087 1088 return ClonedRootL; 1089 } 1090 1091 /// Build the cloned loops of an original loop from unswitching. 1092 /// 1093 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial 1094 /// operation. We need to re-verify that there even is a loop (as the backedge 1095 /// may not have been cloned), and even if there are remaining backedges the 1096 /// backedge set may be different. However, we know that each child loop is 1097 /// undisturbed, we only need to find where to place each child loop within 1098 /// either any parent loop or within a cloned version of the original loop. 1099 /// 1100 /// Because child loops may end up cloned outside of any cloned version of the 1101 /// original loop, multiple cloned sibling loops may be created. All of them 1102 /// are returned so that the newly introduced loop nest roots can be 1103 /// identified. 1104 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks, 1105 const ValueToValueMapTy &VMap, LoopInfo &LI, 1106 SmallVectorImpl<Loop *> &NonChildClonedLoops) { 1107 Loop *ClonedL = nullptr; 1108 1109 auto *OrigPH = OrigL.getLoopPreheader(); 1110 auto *OrigHeader = OrigL.getHeader(); 1111 1112 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH)); 1113 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader)); 1114 1115 // We need to know the loops of the cloned exit blocks to even compute the 1116 // accurate parent loop. If we only clone exits to some parent of the 1117 // original parent, we want to clone into that outer loop. We also keep track 1118 // of the loops that our cloned exit blocks participate in. 1119 Loop *ParentL = nullptr; 1120 SmallVector<BasicBlock *, 4> ClonedExitsInLoops; 1121 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap; 1122 ClonedExitsInLoops.reserve(ExitBlocks.size()); 1123 for (auto *ExitBB : ExitBlocks) 1124 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB))) 1125 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1126 ExitLoopMap[ClonedExitBB] = ExitL; 1127 ClonedExitsInLoops.push_back(ClonedExitBB); 1128 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1129 ParentL = ExitL; 1130 } 1131 assert((!ParentL || ParentL == OrigL.getParentLoop() || 1132 ParentL->contains(OrigL.getParentLoop())) && 1133 "The computed parent loop should always contain (or be) the parent of " 1134 "the original loop."); 1135 1136 // We build the set of blocks dominated by the cloned header from the set of 1137 // cloned blocks out of the original loop. While not all of these will 1138 // necessarily be in the cloned loop, it is enough to establish that they 1139 // aren't in unreachable cycles, etc. 1140 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks; 1141 for (auto *BB : OrigL.blocks()) 1142 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB))) 1143 ClonedLoopBlocks.insert(ClonedBB); 1144 1145 // Rebuild the set of blocks that will end up in the cloned loop. We may have 1146 // skipped cloning some region of this loop which can in turn skip some of 1147 // the backedges so we have to rebuild the blocks in the loop based on the 1148 // backedges that remain after cloning. 1149 SmallVector<BasicBlock *, 16> Worklist; 1150 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop; 1151 for (auto *Pred : predecessors(ClonedHeader)) { 1152 // The only possible non-loop header predecessor is the preheader because 1153 // we know we cloned the loop in simplified form. 1154 if (Pred == ClonedPH) 1155 continue; 1156 1157 // Because the loop was in simplified form, the only non-loop predecessor 1158 // should be the preheader. 1159 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop " 1160 "header other than the preheader " 1161 "that is not part of the loop!"); 1162 1163 // Insert this block into the loop set and on the first visit (and if it 1164 // isn't the header we're currently walking) put it into the worklist to 1165 // recurse through. 1166 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader) 1167 Worklist.push_back(Pred); 1168 } 1169 1170 // If we had any backedges then there *is* a cloned loop. Put the header into 1171 // the loop set and then walk the worklist backwards to find all the blocks 1172 // that remain within the loop after cloning. 1173 if (!BlocksInClonedLoop.empty()) { 1174 BlocksInClonedLoop.insert(ClonedHeader); 1175 1176 while (!Worklist.empty()) { 1177 BasicBlock *BB = Worklist.pop_back_val(); 1178 assert(BlocksInClonedLoop.count(BB) && 1179 "Didn't put block into the loop set!"); 1180 1181 // Insert any predecessors that are in the possible set into the cloned 1182 // set, and if the insert is successful, add them to the worklist. Note 1183 // that we filter on the blocks that are definitely reachable via the 1184 // backedge to the loop header so we may prune out dead code within the 1185 // cloned loop. 1186 for (auto *Pred : predecessors(BB)) 1187 if (ClonedLoopBlocks.count(Pred) && 1188 BlocksInClonedLoop.insert(Pred).second) 1189 Worklist.push_back(Pred); 1190 } 1191 1192 ClonedL = LI.AllocateLoop(); 1193 if (ParentL) { 1194 ParentL->addBasicBlockToLoop(ClonedPH, LI); 1195 ParentL->addChildLoop(ClonedL); 1196 } else { 1197 LI.addTopLevelLoop(ClonedL); 1198 } 1199 NonChildClonedLoops.push_back(ClonedL); 1200 1201 ClonedL->reserveBlocks(BlocksInClonedLoop.size()); 1202 // We don't want to just add the cloned loop blocks based on how we 1203 // discovered them. The original order of blocks was carefully built in 1204 // a way that doesn't rely on predecessor ordering. Rather than re-invent 1205 // that logic, we just re-walk the original blocks (and those of the child 1206 // loops) and filter them as we add them into the cloned loop. 1207 for (auto *BB : OrigL.blocks()) { 1208 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)); 1209 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB)) 1210 continue; 1211 1212 // Directly add the blocks that are only in this loop. 1213 if (LI.getLoopFor(BB) == &OrigL) { 1214 ClonedL->addBasicBlockToLoop(ClonedBB, LI); 1215 continue; 1216 } 1217 1218 // We want to manually add it to this loop and parents. 1219 // Registering it with LoopInfo will happen when we clone the top 1220 // loop for this block. 1221 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop()) 1222 PL->addBlockEntry(ClonedBB); 1223 } 1224 1225 // Now add each child loop whose header remains within the cloned loop. All 1226 // of the blocks within the loop must satisfy the same constraints as the 1227 // header so once we pass the header checks we can just clone the entire 1228 // child loop nest. 1229 for (Loop *ChildL : OrigL) { 1230 auto *ClonedChildHeader = 1231 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1232 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader)) 1233 continue; 1234 1235 #ifndef NDEBUG 1236 // We should never have a cloned child loop header but fail to have 1237 // all of the blocks for that child loop. 1238 for (auto *ChildLoopBB : ChildL->blocks()) 1239 assert(BlocksInClonedLoop.count( 1240 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && 1241 "Child cloned loop has a header within the cloned outer " 1242 "loop but not all of its blocks!"); 1243 #endif 1244 1245 cloneLoopNest(*ChildL, ClonedL, VMap, LI); 1246 } 1247 } 1248 1249 // Now that we've handled all the components of the original loop that were 1250 // cloned into a new loop, we still need to handle anything from the original 1251 // loop that wasn't in a cloned loop. 1252 1253 // Figure out what blocks are left to place within any loop nest containing 1254 // the unswitched loop. If we never formed a loop, the cloned PH is one of 1255 // them. 1256 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet; 1257 if (BlocksInClonedLoop.empty()) 1258 UnloopedBlockSet.insert(ClonedPH); 1259 for (auto *ClonedBB : ClonedLoopBlocks) 1260 if (!BlocksInClonedLoop.count(ClonedBB)) 1261 UnloopedBlockSet.insert(ClonedBB); 1262 1263 // Copy the cloned exits and sort them in ascending loop depth, we'll work 1264 // backwards across these to process them inside out. The order shouldn't 1265 // matter as we're just trying to build up the map from inside-out; we use 1266 // the map in a more stably ordered way below. 1267 auto OrderedClonedExitsInLoops = ClonedExitsInLoops; 1268 llvm::sort(OrderedClonedExitsInLoops.begin(), OrderedClonedExitsInLoops.end(), 1269 [&](BasicBlock *LHS, BasicBlock *RHS) { 1270 return ExitLoopMap.lookup(LHS)->getLoopDepth() < 1271 ExitLoopMap.lookup(RHS)->getLoopDepth(); 1272 }); 1273 1274 // Populate the existing ExitLoopMap with everything reachable from each 1275 // exit, starting from the inner most exit. 1276 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) { 1277 assert(Worklist.empty() && "Didn't clear worklist!"); 1278 1279 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val(); 1280 Loop *ExitL = ExitLoopMap.lookup(ExitBB); 1281 1282 // Walk the CFG back until we hit the cloned PH adding everything reachable 1283 // and in the unlooped set to this exit block's loop. 1284 Worklist.push_back(ExitBB); 1285 do { 1286 BasicBlock *BB = Worklist.pop_back_val(); 1287 // We can stop recursing at the cloned preheader (if we get there). 1288 if (BB == ClonedPH) 1289 continue; 1290 1291 for (BasicBlock *PredBB : predecessors(BB)) { 1292 // If this pred has already been moved to our set or is part of some 1293 // (inner) loop, no update needed. 1294 if (!UnloopedBlockSet.erase(PredBB)) { 1295 assert( 1296 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && 1297 "Predecessor not mapped to a loop!"); 1298 continue; 1299 } 1300 1301 // We just insert into the loop set here. We'll add these blocks to the 1302 // exit loop after we build up the set in an order that doesn't rely on 1303 // predecessor order (which in turn relies on use list order). 1304 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second; 1305 (void)Inserted; 1306 assert(Inserted && "Should only visit an unlooped block once!"); 1307 1308 // And recurse through to its predecessors. 1309 Worklist.push_back(PredBB); 1310 } 1311 } while (!Worklist.empty()); 1312 } 1313 1314 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned 1315 // blocks to their outer loops, walk the cloned blocks and the cloned exits 1316 // in their original order adding them to the correct loop. 1317 1318 // We need a stable insertion order. We use the order of the original loop 1319 // order and map into the correct parent loop. 1320 for (auto *BB : llvm::concat<BasicBlock *const>( 1321 makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops)) 1322 if (Loop *OuterL = ExitLoopMap.lookup(BB)) 1323 OuterL->addBasicBlockToLoop(BB, LI); 1324 1325 #ifndef NDEBUG 1326 for (auto &BBAndL : ExitLoopMap) { 1327 auto *BB = BBAndL.first; 1328 auto *OuterL = BBAndL.second; 1329 assert(LI.getLoopFor(BB) == OuterL && 1330 "Failed to put all blocks into outer loops!"); 1331 } 1332 #endif 1333 1334 // Now that all the blocks are placed into the correct containing loop in the 1335 // absence of child loops, find all the potentially cloned child loops and 1336 // clone them into whatever outer loop we placed their header into. 1337 for (Loop *ChildL : OrigL) { 1338 auto *ClonedChildHeader = 1339 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1340 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader)) 1341 continue; 1342 1343 #ifndef NDEBUG 1344 for (auto *ChildLoopBB : ChildL->blocks()) 1345 assert(VMap.count(ChildLoopBB) && 1346 "Cloned a child loop header but not all of that loops blocks!"); 1347 #endif 1348 1349 NonChildClonedLoops.push_back(cloneLoopNest( 1350 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI)); 1351 } 1352 } 1353 1354 static void 1355 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1356 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, 1357 DominatorTree &DT) { 1358 // Find all the dead clones, and remove them from their successors. 1359 SmallVector<BasicBlock *, 16> DeadBlocks; 1360 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks)) 1361 for (auto &VMap : VMaps) 1362 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB))) 1363 if (!DT.isReachableFromEntry(ClonedBB)) { 1364 for (BasicBlock *SuccBB : successors(ClonedBB)) 1365 SuccBB->removePredecessor(ClonedBB); 1366 DeadBlocks.push_back(ClonedBB); 1367 } 1368 1369 // Drop any remaining references to break cycles. 1370 for (BasicBlock *BB : DeadBlocks) 1371 BB->dropAllReferences(); 1372 // Erase them from the IR. 1373 for (BasicBlock *BB : DeadBlocks) 1374 BB->eraseFromParent(); 1375 } 1376 1377 static void 1378 deleteDeadBlocksFromLoop(Loop &L, 1379 SmallVectorImpl<BasicBlock *> &ExitBlocks, 1380 DominatorTree &DT, LoopInfo &LI) { 1381 // Find all the dead blocks tied to this loop, and remove them from their 1382 // successors. 1383 SmallPtrSet<BasicBlock *, 16> DeadBlockSet; 1384 1385 // Start with loop/exit blocks and get a transitive closure of reachable dead 1386 // blocks. 1387 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(), 1388 ExitBlocks.end()); 1389 DeathCandidates.append(L.blocks().begin(), L.blocks().end()); 1390 while (!DeathCandidates.empty()) { 1391 auto *BB = DeathCandidates.pop_back_val(); 1392 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) { 1393 for (BasicBlock *SuccBB : successors(BB)) { 1394 SuccBB->removePredecessor(BB); 1395 DeathCandidates.push_back(SuccBB); 1396 } 1397 DeadBlockSet.insert(BB); 1398 } 1399 } 1400 1401 // Filter out the dead blocks from the exit blocks list so that it can be 1402 // used in the caller. 1403 llvm::erase_if(ExitBlocks, 1404 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1405 1406 // Walk from this loop up through its parents removing all of the dead blocks. 1407 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) { 1408 for (auto *BB : DeadBlockSet) 1409 ParentL->getBlocksSet().erase(BB); 1410 llvm::erase_if(ParentL->getBlocksVector(), 1411 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1412 } 1413 1414 // Now delete the dead child loops. This raw delete will clear them 1415 // recursively. 1416 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) { 1417 if (!DeadBlockSet.count(ChildL->getHeader())) 1418 return false; 1419 1420 assert(llvm::all_of(ChildL->blocks(), 1421 [&](BasicBlock *ChildBB) { 1422 return DeadBlockSet.count(ChildBB); 1423 }) && 1424 "If the child loop header is dead all blocks in the child loop must " 1425 "be dead as well!"); 1426 LI.destroy(ChildL); 1427 return true; 1428 }); 1429 1430 // Remove the loop mappings for the dead blocks and drop all the references 1431 // from these blocks to others to handle cyclic references as we start 1432 // deleting the blocks themselves. 1433 for (auto *BB : DeadBlockSet) { 1434 // Check that the dominator tree has already been updated. 1435 assert(!DT.getNode(BB) && "Should already have cleared domtree!"); 1436 LI.changeLoopFor(BB, nullptr); 1437 BB->dropAllReferences(); 1438 } 1439 1440 // Actually delete the blocks now that they've been fully unhooked from the 1441 // IR. 1442 for (auto *BB : DeadBlockSet) 1443 BB->eraseFromParent(); 1444 } 1445 1446 /// Recompute the set of blocks in a loop after unswitching. 1447 /// 1448 /// This walks from the original headers predecessors to rebuild the loop. We 1449 /// take advantage of the fact that new blocks can't have been added, and so we 1450 /// filter by the original loop's blocks. This also handles potentially 1451 /// unreachable code that we don't want to explore but might be found examining 1452 /// the predecessors of the header. 1453 /// 1454 /// If the original loop is no longer a loop, this will return an empty set. If 1455 /// it remains a loop, all the blocks within it will be added to the set 1456 /// (including those blocks in inner loops). 1457 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L, 1458 LoopInfo &LI) { 1459 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet; 1460 1461 auto *PH = L.getLoopPreheader(); 1462 auto *Header = L.getHeader(); 1463 1464 // A worklist to use while walking backwards from the header. 1465 SmallVector<BasicBlock *, 16> Worklist; 1466 1467 // First walk the predecessors of the header to find the backedges. This will 1468 // form the basis of our walk. 1469 for (auto *Pred : predecessors(Header)) { 1470 // Skip the preheader. 1471 if (Pred == PH) 1472 continue; 1473 1474 // Because the loop was in simplified form, the only non-loop predecessor 1475 // is the preheader. 1476 assert(L.contains(Pred) && "Found a predecessor of the loop header other " 1477 "than the preheader that is not part of the " 1478 "loop!"); 1479 1480 // Insert this block into the loop set and on the first visit and, if it 1481 // isn't the header we're currently walking, put it into the worklist to 1482 // recurse through. 1483 if (LoopBlockSet.insert(Pred).second && Pred != Header) 1484 Worklist.push_back(Pred); 1485 } 1486 1487 // If no backedges were found, we're done. 1488 if (LoopBlockSet.empty()) 1489 return LoopBlockSet; 1490 1491 // We found backedges, recurse through them to identify the loop blocks. 1492 while (!Worklist.empty()) { 1493 BasicBlock *BB = Worklist.pop_back_val(); 1494 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!"); 1495 1496 // No need to walk past the header. 1497 if (BB == Header) 1498 continue; 1499 1500 // Because we know the inner loop structure remains valid we can use the 1501 // loop structure to jump immediately across the entire nested loop. 1502 // Further, because it is in loop simplified form, we can directly jump 1503 // to its preheader afterward. 1504 if (Loop *InnerL = LI.getLoopFor(BB)) 1505 if (InnerL != &L) { 1506 assert(L.contains(InnerL) && 1507 "Should not reach a loop *outside* this loop!"); 1508 // The preheader is the only possible predecessor of the loop so 1509 // insert it into the set and check whether it was already handled. 1510 auto *InnerPH = InnerL->getLoopPreheader(); 1511 assert(L.contains(InnerPH) && "Cannot contain an inner loop block " 1512 "but not contain the inner loop " 1513 "preheader!"); 1514 if (!LoopBlockSet.insert(InnerPH).second) 1515 // The only way to reach the preheader is through the loop body 1516 // itself so if it has been visited the loop is already handled. 1517 continue; 1518 1519 // Insert all of the blocks (other than those already present) into 1520 // the loop set. We expect at least the block that led us to find the 1521 // inner loop to be in the block set, but we may also have other loop 1522 // blocks if they were already enqueued as predecessors of some other 1523 // outer loop block. 1524 for (auto *InnerBB : InnerL->blocks()) { 1525 if (InnerBB == BB) { 1526 assert(LoopBlockSet.count(InnerBB) && 1527 "Block should already be in the set!"); 1528 continue; 1529 } 1530 1531 LoopBlockSet.insert(InnerBB); 1532 } 1533 1534 // Add the preheader to the worklist so we will continue past the 1535 // loop body. 1536 Worklist.push_back(InnerPH); 1537 continue; 1538 } 1539 1540 // Insert any predecessors that were in the original loop into the new 1541 // set, and if the insert is successful, add them to the worklist. 1542 for (auto *Pred : predecessors(BB)) 1543 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second) 1544 Worklist.push_back(Pred); 1545 } 1546 1547 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!"); 1548 1549 // We've found all the blocks participating in the loop, return our completed 1550 // set. 1551 return LoopBlockSet; 1552 } 1553 1554 /// Rebuild a loop after unswitching removes some subset of blocks and edges. 1555 /// 1556 /// The removal may have removed some child loops entirely but cannot have 1557 /// disturbed any remaining child loops. However, they may need to be hoisted 1558 /// to the parent loop (or to be top-level loops). The original loop may be 1559 /// completely removed. 1560 /// 1561 /// The sibling loops resulting from this update are returned. If the original 1562 /// loop remains a valid loop, it will be the first entry in this list with all 1563 /// of the newly sibling loops following it. 1564 /// 1565 /// Returns true if the loop remains a loop after unswitching, and false if it 1566 /// is no longer a loop after unswitching (and should not continue to be 1567 /// referenced). 1568 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1569 LoopInfo &LI, 1570 SmallVectorImpl<Loop *> &HoistedLoops) { 1571 auto *PH = L.getLoopPreheader(); 1572 1573 // Compute the actual parent loop from the exit blocks. Because we may have 1574 // pruned some exits the loop may be different from the original parent. 1575 Loop *ParentL = nullptr; 1576 SmallVector<Loop *, 4> ExitLoops; 1577 SmallVector<BasicBlock *, 4> ExitsInLoops; 1578 ExitsInLoops.reserve(ExitBlocks.size()); 1579 for (auto *ExitBB : ExitBlocks) 1580 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1581 ExitLoops.push_back(ExitL); 1582 ExitsInLoops.push_back(ExitBB); 1583 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1584 ParentL = ExitL; 1585 } 1586 1587 // Recompute the blocks participating in this loop. This may be empty if it 1588 // is no longer a loop. 1589 auto LoopBlockSet = recomputeLoopBlockSet(L, LI); 1590 1591 // If we still have a loop, we need to re-set the loop's parent as the exit 1592 // block set changing may have moved it within the loop nest. Note that this 1593 // can only happen when this loop has a parent as it can only hoist the loop 1594 // *up* the nest. 1595 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) { 1596 // Remove this loop's (original) blocks from all of the intervening loops. 1597 for (Loop *IL = L.getParentLoop(); IL != ParentL; 1598 IL = IL->getParentLoop()) { 1599 IL->getBlocksSet().erase(PH); 1600 for (auto *BB : L.blocks()) 1601 IL->getBlocksSet().erase(BB); 1602 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) { 1603 return BB == PH || L.contains(BB); 1604 }); 1605 } 1606 1607 LI.changeLoopFor(PH, ParentL); 1608 L.getParentLoop()->removeChildLoop(&L); 1609 if (ParentL) 1610 ParentL->addChildLoop(&L); 1611 else 1612 LI.addTopLevelLoop(&L); 1613 } 1614 1615 // Now we update all the blocks which are no longer within the loop. 1616 auto &Blocks = L.getBlocksVector(); 1617 auto BlocksSplitI = 1618 LoopBlockSet.empty() 1619 ? Blocks.begin() 1620 : std::stable_partition( 1621 Blocks.begin(), Blocks.end(), 1622 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); }); 1623 1624 // Before we erase the list of unlooped blocks, build a set of them. 1625 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end()); 1626 if (LoopBlockSet.empty()) 1627 UnloopedBlocks.insert(PH); 1628 1629 // Now erase these blocks from the loop. 1630 for (auto *BB : make_range(BlocksSplitI, Blocks.end())) 1631 L.getBlocksSet().erase(BB); 1632 Blocks.erase(BlocksSplitI, Blocks.end()); 1633 1634 // Sort the exits in ascending loop depth, we'll work backwards across these 1635 // to process them inside out. 1636 std::stable_sort(ExitsInLoops.begin(), ExitsInLoops.end(), 1637 [&](BasicBlock *LHS, BasicBlock *RHS) { 1638 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS); 1639 }); 1640 1641 // We'll build up a set for each exit loop. 1642 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks; 1643 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop. 1644 1645 auto RemoveUnloopedBlocksFromLoop = 1646 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) { 1647 for (auto *BB : UnloopedBlocks) 1648 L.getBlocksSet().erase(BB); 1649 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) { 1650 return UnloopedBlocks.count(BB); 1651 }); 1652 }; 1653 1654 SmallVector<BasicBlock *, 16> Worklist; 1655 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) { 1656 assert(Worklist.empty() && "Didn't clear worklist!"); 1657 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!"); 1658 1659 // Grab the next exit block, in decreasing loop depth order. 1660 BasicBlock *ExitBB = ExitsInLoops.pop_back_val(); 1661 Loop &ExitL = *LI.getLoopFor(ExitBB); 1662 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!"); 1663 1664 // Erase all of the unlooped blocks from the loops between the previous 1665 // exit loop and this exit loop. This works because the ExitInLoops list is 1666 // sorted in increasing order of loop depth and thus we visit loops in 1667 // decreasing order of loop depth. 1668 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop()) 1669 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1670 1671 // Walk the CFG back until we hit the cloned PH adding everything reachable 1672 // and in the unlooped set to this exit block's loop. 1673 Worklist.push_back(ExitBB); 1674 do { 1675 BasicBlock *BB = Worklist.pop_back_val(); 1676 // We can stop recursing at the cloned preheader (if we get there). 1677 if (BB == PH) 1678 continue; 1679 1680 for (BasicBlock *PredBB : predecessors(BB)) { 1681 // If this pred has already been moved to our set or is part of some 1682 // (inner) loop, no update needed. 1683 if (!UnloopedBlocks.erase(PredBB)) { 1684 assert((NewExitLoopBlocks.count(PredBB) || 1685 ExitL.contains(LI.getLoopFor(PredBB))) && 1686 "Predecessor not in a nested loop (or already visited)!"); 1687 continue; 1688 } 1689 1690 // We just insert into the loop set here. We'll add these blocks to the 1691 // exit loop after we build up the set in a deterministic order rather 1692 // than the predecessor-influenced visit order. 1693 bool Inserted = NewExitLoopBlocks.insert(PredBB).second; 1694 (void)Inserted; 1695 assert(Inserted && "Should only visit an unlooped block once!"); 1696 1697 // And recurse through to its predecessors. 1698 Worklist.push_back(PredBB); 1699 } 1700 } while (!Worklist.empty()); 1701 1702 // If blocks in this exit loop were directly part of the original loop (as 1703 // opposed to a child loop) update the map to point to this exit loop. This 1704 // just updates a map and so the fact that the order is unstable is fine. 1705 for (auto *BB : NewExitLoopBlocks) 1706 if (Loop *BBL = LI.getLoopFor(BB)) 1707 if (BBL == &L || !L.contains(BBL)) 1708 LI.changeLoopFor(BB, &ExitL); 1709 1710 // We will remove the remaining unlooped blocks from this loop in the next 1711 // iteration or below. 1712 NewExitLoopBlocks.clear(); 1713 } 1714 1715 // Any remaining unlooped blocks are no longer part of any loop unless they 1716 // are part of some child loop. 1717 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop()) 1718 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1719 for (auto *BB : UnloopedBlocks) 1720 if (Loop *BBL = LI.getLoopFor(BB)) 1721 if (BBL == &L || !L.contains(BBL)) 1722 LI.changeLoopFor(BB, nullptr); 1723 1724 // Sink all the child loops whose headers are no longer in the loop set to 1725 // the parent (or to be top level loops). We reach into the loop and directly 1726 // update its subloop vector to make this batch update efficient. 1727 auto &SubLoops = L.getSubLoopsVector(); 1728 auto SubLoopsSplitI = 1729 LoopBlockSet.empty() 1730 ? SubLoops.begin() 1731 : std::stable_partition( 1732 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) { 1733 return LoopBlockSet.count(SubL->getHeader()); 1734 }); 1735 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) { 1736 HoistedLoops.push_back(HoistedL); 1737 HoistedL->setParentLoop(nullptr); 1738 1739 // To compute the new parent of this hoisted loop we look at where we 1740 // placed the preheader above. We can't lookup the header itself because we 1741 // retained the mapping from the header to the hoisted loop. But the 1742 // preheader and header should have the exact same new parent computed 1743 // based on the set of exit blocks from the original loop as the preheader 1744 // is a predecessor of the header and so reached in the reverse walk. And 1745 // because the loops were all in simplified form the preheader of the 1746 // hoisted loop can't be part of some *other* loop. 1747 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader())) 1748 NewParentL->addChildLoop(HoistedL); 1749 else 1750 LI.addTopLevelLoop(HoistedL); 1751 } 1752 SubLoops.erase(SubLoopsSplitI, SubLoops.end()); 1753 1754 // Actually delete the loop if nothing remained within it. 1755 if (Blocks.empty()) { 1756 assert(SubLoops.empty() && 1757 "Failed to remove all subloops from the original loop!"); 1758 if (Loop *ParentL = L.getParentLoop()) 1759 ParentL->removeChildLoop(llvm::find(*ParentL, &L)); 1760 else 1761 LI.removeLoop(llvm::find(LI, &L)); 1762 LI.destroy(&L); 1763 return false; 1764 } 1765 1766 return true; 1767 } 1768 1769 /// Helper to visit a dominator subtree, invoking a callable on each node. 1770 /// 1771 /// Returning false at any point will stop walking past that node of the tree. 1772 template <typename CallableT> 1773 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) { 1774 SmallVector<DomTreeNode *, 4> DomWorklist; 1775 DomWorklist.push_back(DT[BB]); 1776 #ifndef NDEBUG 1777 SmallPtrSet<DomTreeNode *, 4> Visited; 1778 Visited.insert(DT[BB]); 1779 #endif 1780 do { 1781 DomTreeNode *N = DomWorklist.pop_back_val(); 1782 1783 // Visit this node. 1784 if (!Callable(N->getBlock())) 1785 continue; 1786 1787 // Accumulate the child nodes. 1788 for (DomTreeNode *ChildN : *N) { 1789 assert(Visited.insert(ChildN).second && 1790 "Cannot visit a node twice when walking a tree!"); 1791 DomWorklist.push_back(ChildN); 1792 } 1793 } while (!DomWorklist.empty()); 1794 } 1795 1796 static bool unswitchNontrivialInvariants( 1797 Loop &L, TerminatorInst &TI, ArrayRef<Value *> Invariants, 1798 DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, 1799 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB, 1800 ScalarEvolution *SE) { 1801 auto *ParentBB = TI.getParent(); 1802 BranchInst *BI = dyn_cast<BranchInst>(&TI); 1803 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI); 1804 1805 // We can only unswitch switches, conditional branches with an invariant 1806 // condition, or combining invariant conditions with an instruction. 1807 assert((SI || BI->isConditional()) && 1808 "Can only unswitch switches and conditional branch!"); 1809 bool FullUnswitch = SI || BI->getCondition() == Invariants[0]; 1810 if (FullUnswitch) 1811 assert(Invariants.size() == 1 && 1812 "Cannot have other invariants with full unswitching!"); 1813 else 1814 assert(isa<Instruction>(BI->getCondition()) && 1815 "Partial unswitching requires an instruction as the condition!"); 1816 1817 // Constant and BBs tracking the cloned and continuing successor. When we are 1818 // unswitching the entire condition, this can just be trivially chosen to 1819 // unswitch towards `true`. However, when we are unswitching a set of 1820 // invariants combined with `and` or `or`, the combining operation determines 1821 // the best direction to unswitch: we want to unswitch the direction that will 1822 // collapse the branch. 1823 bool Direction = true; 1824 int ClonedSucc = 0; 1825 if (!FullUnswitch) { 1826 if (cast<Instruction>(BI->getCondition())->getOpcode() != Instruction::Or) { 1827 assert(cast<Instruction>(BI->getCondition())->getOpcode() == 1828 Instruction::And && 1829 "Only `or` and `and` instructions can combine invariants being " 1830 "unswitched."); 1831 Direction = false; 1832 ClonedSucc = 1; 1833 } 1834 } 1835 1836 BasicBlock *RetainedSuccBB = 1837 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest(); 1838 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs; 1839 if (BI) 1840 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc)); 1841 else 1842 for (auto Case : SI->cases()) 1843 if (Case.getCaseSuccessor() != RetainedSuccBB) 1844 UnswitchedSuccBBs.insert(Case.getCaseSuccessor()); 1845 1846 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) && 1847 "Should not unswitch the same successor we are retaining!"); 1848 1849 // The branch should be in this exact loop. Any inner loop's invariant branch 1850 // should be handled by unswitching that inner loop. The caller of this 1851 // routine should filter out any candidates that remain (but were skipped for 1852 // whatever reason). 1853 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!"); 1854 1855 SmallVector<BasicBlock *, 4> ExitBlocks; 1856 L.getUniqueExitBlocks(ExitBlocks); 1857 1858 // We cannot unswitch if exit blocks contain a cleanuppad instruction as we 1859 // don't know how to split those exit blocks. 1860 // FIXME: We should teach SplitBlock to handle this and remove this 1861 // restriction. 1862 for (auto *ExitBB : ExitBlocks) 1863 if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) 1864 return false; 1865 1866 // Compute the parent loop now before we start hacking on things. 1867 Loop *ParentL = L.getParentLoop(); 1868 1869 // Compute the outer-most loop containing one of our exit blocks. This is the 1870 // furthest up our loopnest which can be mutated, which we will use below to 1871 // update things. 1872 Loop *OuterExitL = &L; 1873 for (auto *ExitBB : ExitBlocks) { 1874 Loop *NewOuterExitL = LI.getLoopFor(ExitBB); 1875 if (!NewOuterExitL) { 1876 // We exited the entire nest with this block, so we're done. 1877 OuterExitL = nullptr; 1878 break; 1879 } 1880 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL)) 1881 OuterExitL = NewOuterExitL; 1882 } 1883 1884 // At this point, we're definitely going to unswitch something so invalidate 1885 // any cached information in ScalarEvolution for the outer most loop 1886 // containing an exit block and all nested loops. 1887 if (SE) { 1888 if (OuterExitL) 1889 SE->forgetLoop(OuterExitL); 1890 else 1891 SE->forgetTopmostLoop(&L); 1892 } 1893 1894 // If the edge from this terminator to a successor dominates that successor, 1895 // store a map from each block in its dominator subtree to it. This lets us 1896 // tell when cloning for a particular successor if a block is dominated by 1897 // some *other* successor with a single data structure. We use this to 1898 // significantly reduce cloning. 1899 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc; 1900 for (auto *SuccBB : llvm::concat<BasicBlock *const>( 1901 makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs)) 1902 if (SuccBB->getUniquePredecessor() || 1903 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 1904 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB); 1905 })) 1906 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) { 1907 DominatingSucc[BB] = SuccBB; 1908 return true; 1909 }); 1910 1911 // Split the preheader, so that we know that there is a safe place to insert 1912 // the conditional branch. We will change the preheader to have a conditional 1913 // branch on LoopCond. The original preheader will become the split point 1914 // between the unswitched versions, and we will have a new preheader for the 1915 // original loop. 1916 BasicBlock *SplitBB = L.getLoopPreheader(); 1917 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI); 1918 1919 // Keep track of the dominator tree updates needed. 1920 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 1921 1922 // Clone the loop for each unswitched successor. 1923 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; 1924 VMaps.reserve(UnswitchedSuccBBs.size()); 1925 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs; 1926 for (auto *SuccBB : UnswitchedSuccBBs) { 1927 VMaps.emplace_back(new ValueToValueMapTy()); 1928 ClonedPHs[SuccBB] = buildClonedLoopBlocks( 1929 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB, 1930 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI); 1931 } 1932 1933 // The stitching of the branched code back together depends on whether we're 1934 // doing full unswitching or not with the exception that we always want to 1935 // nuke the initial terminator placed in the split block. 1936 SplitBB->getTerminator()->eraseFromParent(); 1937 if (FullUnswitch) { 1938 // First we need to unhook the successor relationship as we'll be replacing 1939 // the terminator with a direct branch. This is much simpler for branches 1940 // than switches so we handle those first. 1941 if (BI) { 1942 // Remove the parent as a predecessor of the unswitched successor. 1943 assert(UnswitchedSuccBBs.size() == 1 && 1944 "Only one possible unswitched block for a branch!"); 1945 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin(); 1946 UnswitchedSuccBB->removePredecessor(ParentBB, 1947 /*DontDeleteUselessPHIs*/ true); 1948 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB}); 1949 } else { 1950 // Note that we actually want to remove the parent block as a predecessor 1951 // of *every* case successor. The case successor is either unswitched, 1952 // completely eliminating an edge from the parent to that successor, or it 1953 // is a duplicate edge to the retained successor as the retained successor 1954 // is always the default successor and as we'll replace this with a direct 1955 // branch we no longer need the duplicate entries in the PHI nodes. 1956 assert(SI->getDefaultDest() == RetainedSuccBB && 1957 "Not retaining default successor!"); 1958 for (auto &Case : SI->cases()) 1959 Case.getCaseSuccessor()->removePredecessor( 1960 ParentBB, 1961 /*DontDeleteUselessPHIs*/ true); 1962 1963 // We need to use the set to populate domtree updates as even when there 1964 // are multiple cases pointing at the same successor we only want to 1965 // remove and insert one edge in the domtree. 1966 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 1967 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB}); 1968 } 1969 1970 // Now that we've unhooked the successor relationship, splice the terminator 1971 // from the original loop to the split. 1972 SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI); 1973 1974 // Now wire up the terminator to the preheaders. 1975 if (BI) { 1976 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 1977 BI->setSuccessor(ClonedSucc, ClonedPH); 1978 BI->setSuccessor(1 - ClonedSucc, LoopPH); 1979 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 1980 } else { 1981 assert(SI && "Must either be a branch or switch!"); 1982 1983 // Walk the cases and directly update their successors. 1984 SI->setDefaultDest(LoopPH); 1985 for (auto &Case : SI->cases()) 1986 if (Case.getCaseSuccessor() == RetainedSuccBB) 1987 Case.setSuccessor(LoopPH); 1988 else 1989 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second); 1990 1991 // We need to use the set to populate domtree updates as even when there 1992 // are multiple cases pointing at the same successor we only want to 1993 // remove and insert one edge in the domtree. 1994 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 1995 DTUpdates.push_back( 1996 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second}); 1997 } 1998 1999 // Create a new unconditional branch to the continuing block (as opposed to 2000 // the one cloned). 2001 BranchInst::Create(RetainedSuccBB, ParentBB); 2002 } else { 2003 assert(BI && "Only branches have partial unswitching."); 2004 assert(UnswitchedSuccBBs.size() == 1 && 2005 "Only one possible unswitched block for a branch!"); 2006 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2007 // When doing a partial unswitch, we have to do a bit more work to build up 2008 // the branch in the split block. 2009 buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction, 2010 *ClonedPH, *LoopPH); 2011 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2012 } 2013 2014 // Apply the updates accumulated above to get an up-to-date dominator tree. 2015 DT.applyUpdates(DTUpdates); 2016 2017 // Now that we have an accurate dominator tree, first delete the dead cloned 2018 // blocks so that we can accurately build any cloned loops. It is important to 2019 // not delete the blocks from the original loop yet because we still want to 2020 // reference the original loop to understand the cloned loop's structure. 2021 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT); 2022 2023 // Build the cloned loop structure itself. This may be substantially 2024 // different from the original structure due to the simplified CFG. This also 2025 // handles inserting all the cloned blocks into the correct loops. 2026 SmallVector<Loop *, 4> NonChildClonedLoops; 2027 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps) 2028 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops); 2029 2030 // Now that our cloned loops have been built, we can update the original loop. 2031 // First we delete the dead blocks from it and then we rebuild the loop 2032 // structure taking these deletions into account. 2033 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI); 2034 SmallVector<Loop *, 4> HoistedLoops; 2035 bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops); 2036 2037 // This transformation has a high risk of corrupting the dominator tree, and 2038 // the below steps to rebuild loop structures will result in hard to debug 2039 // errors in that case so verify that the dominator tree is sane first. 2040 // FIXME: Remove this when the bugs stop showing up and rely on existing 2041 // verification steps. 2042 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2043 2044 if (BI) { 2045 // If we unswitched a branch which collapses the condition to a known 2046 // constant we want to replace all the uses of the invariants within both 2047 // the original and cloned blocks. We do this here so that we can use the 2048 // now updated dominator tree to identify which side the users are on. 2049 assert(UnswitchedSuccBBs.size() == 1 && 2050 "Only one possible unswitched block for a branch!"); 2051 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2052 ConstantInt *UnswitchedReplacement = 2053 Direction ? ConstantInt::getTrue(BI->getContext()) 2054 : ConstantInt::getFalse(BI->getContext()); 2055 ConstantInt *ContinueReplacement = 2056 Direction ? ConstantInt::getFalse(BI->getContext()) 2057 : ConstantInt::getTrue(BI->getContext()); 2058 for (Value *Invariant : Invariants) 2059 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); 2060 UI != UE;) { 2061 // Grab the use and walk past it so we can clobber it in the use list. 2062 Use *U = &*UI++; 2063 Instruction *UserI = dyn_cast<Instruction>(U->getUser()); 2064 if (!UserI) 2065 continue; 2066 2067 // Replace it with the 'continue' side if in the main loop body, and the 2068 // unswitched if in the cloned blocks. 2069 if (DT.dominates(LoopPH, UserI->getParent())) 2070 U->set(ContinueReplacement); 2071 else if (DT.dominates(ClonedPH, UserI->getParent())) 2072 U->set(UnswitchedReplacement); 2073 } 2074 } 2075 2076 // We can change which blocks are exit blocks of all the cloned sibling 2077 // loops, the current loop, and any parent loops which shared exit blocks 2078 // with the current loop. As a consequence, we need to re-form LCSSA for 2079 // them. But we shouldn't need to re-form LCSSA for any child loops. 2080 // FIXME: This could be made more efficient by tracking which exit blocks are 2081 // new, and focusing on them, but that isn't likely to be necessary. 2082 // 2083 // In order to reasonably rebuild LCSSA we need to walk inside-out across the 2084 // loop nest and update every loop that could have had its exits changed. We 2085 // also need to cover any intervening loops. We add all of these loops to 2086 // a list and sort them by loop depth to achieve this without updating 2087 // unnecessary loops. 2088 auto UpdateLoop = [&](Loop &UpdateL) { 2089 #ifndef NDEBUG 2090 UpdateL.verifyLoop(); 2091 for (Loop *ChildL : UpdateL) { 2092 ChildL->verifyLoop(); 2093 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) && 2094 "Perturbed a child loop's LCSSA form!"); 2095 } 2096 #endif 2097 // First build LCSSA for this loop so that we can preserve it when 2098 // forming dedicated exits. We don't want to perturb some other loop's 2099 // LCSSA while doing that CFG edit. 2100 formLCSSA(UpdateL, DT, &LI, nullptr); 2101 2102 // For loops reached by this loop's original exit blocks we may 2103 // introduced new, non-dedicated exits. At least try to re-form dedicated 2104 // exits for these loops. This may fail if they couldn't have dedicated 2105 // exits to start with. 2106 formDedicatedExitBlocks(&UpdateL, &DT, &LI, /*PreserveLCSSA*/ true); 2107 }; 2108 2109 // For non-child cloned loops and hoisted loops, we just need to update LCSSA 2110 // and we can do it in any order as they don't nest relative to each other. 2111 // 2112 // Also check if any of the loops we have updated have become top-level loops 2113 // as that will necessitate widening the outer loop scope. 2114 for (Loop *UpdatedL : 2115 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) { 2116 UpdateLoop(*UpdatedL); 2117 if (!UpdatedL->getParentLoop()) 2118 OuterExitL = nullptr; 2119 } 2120 if (IsStillLoop) { 2121 UpdateLoop(L); 2122 if (!L.getParentLoop()) 2123 OuterExitL = nullptr; 2124 } 2125 2126 // If the original loop had exit blocks, walk up through the outer most loop 2127 // of those exit blocks to update LCSSA and form updated dedicated exits. 2128 if (OuterExitL != &L) 2129 for (Loop *OuterL = ParentL; OuterL != OuterExitL; 2130 OuterL = OuterL->getParentLoop()) 2131 UpdateLoop(*OuterL); 2132 2133 #ifndef NDEBUG 2134 // Verify the entire loop structure to catch any incorrect updates before we 2135 // progress in the pass pipeline. 2136 LI.verify(DT); 2137 #endif 2138 2139 // Now that we've unswitched something, make callbacks to report the changes. 2140 // For that we need to merge together the updated loops and the cloned loops 2141 // and check whether the original loop survived. 2142 SmallVector<Loop *, 4> SibLoops; 2143 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) 2144 if (UpdatedL->getParentLoop() == ParentL) 2145 SibLoops.push_back(UpdatedL); 2146 UnswitchCB(IsStillLoop, SibLoops); 2147 2148 ++NumBranches; 2149 return true; 2150 } 2151 2152 /// Recursively compute the cost of a dominator subtree based on the per-block 2153 /// cost map provided. 2154 /// 2155 /// The recursive computation is memozied into the provided DT-indexed cost map 2156 /// to allow querying it for most nodes in the domtree without it becoming 2157 /// quadratic. 2158 static int 2159 computeDomSubtreeCost(DomTreeNode &N, 2160 const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap, 2161 SmallDenseMap<DomTreeNode *, int, 4> &DTCostMap) { 2162 // Don't accumulate cost (or recurse through) blocks not in our block cost 2163 // map and thus not part of the duplication cost being considered. 2164 auto BBCostIt = BBCostMap.find(N.getBlock()); 2165 if (BBCostIt == BBCostMap.end()) 2166 return 0; 2167 2168 // Lookup this node to see if we already computed its cost. 2169 auto DTCostIt = DTCostMap.find(&N); 2170 if (DTCostIt != DTCostMap.end()) 2171 return DTCostIt->second; 2172 2173 // If not, we have to compute it. We can't use insert above and update 2174 // because computing the cost may insert more things into the map. 2175 int Cost = std::accumulate( 2176 N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) { 2177 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap); 2178 }); 2179 bool Inserted = DTCostMap.insert({&N, Cost}).second; 2180 (void)Inserted; 2181 assert(Inserted && "Should not insert a node while visiting children!"); 2182 return Cost; 2183 } 2184 2185 static bool 2186 unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI, 2187 AssumptionCache &AC, TargetTransformInfo &TTI, 2188 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB, 2189 ScalarEvolution *SE) { 2190 // Collect all invariant conditions within this loop (as opposed to an inner 2191 // loop which would be handled when visiting that inner loop). 2192 SmallVector<std::pair<TerminatorInst *, TinyPtrVector<Value *>>, 4> 2193 UnswitchCandidates; 2194 for (auto *BB : L.blocks()) { 2195 if (LI.getLoopFor(BB) != &L) 2196 continue; 2197 2198 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2199 // We can only consider fully loop-invariant switch conditions as we need 2200 // to completely eliminate the switch after unswitching. 2201 if (!isa<Constant>(SI->getCondition()) && 2202 L.isLoopInvariant(SI->getCondition())) 2203 UnswitchCandidates.push_back({SI, {SI->getCondition()}}); 2204 continue; 2205 } 2206 2207 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 2208 if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) || 2209 BI->getSuccessor(0) == BI->getSuccessor(1)) 2210 continue; 2211 2212 if (L.isLoopInvariant(BI->getCondition())) { 2213 UnswitchCandidates.push_back({BI, {BI->getCondition()}}); 2214 continue; 2215 } 2216 2217 Instruction &CondI = *cast<Instruction>(BI->getCondition()); 2218 if (CondI.getOpcode() != Instruction::And && 2219 CondI.getOpcode() != Instruction::Or) 2220 continue; 2221 2222 TinyPtrVector<Value *> Invariants = 2223 collectHomogenousInstGraphLoopInvariants(L, CondI, LI); 2224 if (Invariants.empty()) 2225 continue; 2226 2227 UnswitchCandidates.push_back({BI, std::move(Invariants)}); 2228 } 2229 2230 // If we didn't find any candidates, we're done. 2231 if (UnswitchCandidates.empty()) 2232 return false; 2233 2234 // Check if there are irreducible CFG cycles in this loop. If so, we cannot 2235 // easily unswitch non-trivial edges out of the loop. Doing so might turn the 2236 // irreducible control flow into reducible control flow and introduce new 2237 // loops "out of thin air". If we ever discover important use cases for doing 2238 // this, we can add support to loop unswitch, but it is a lot of complexity 2239 // for what seems little or no real world benefit. 2240 LoopBlocksRPO RPOT(&L); 2241 RPOT.perform(&LI); 2242 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 2243 return false; 2244 2245 LLVM_DEBUG( 2246 dbgs() << "Considering " << UnswitchCandidates.size() 2247 << " non-trivial loop invariant conditions for unswitching.\n"); 2248 2249 // Given that unswitching these terminators will require duplicating parts of 2250 // the loop, so we need to be able to model that cost. Compute the ephemeral 2251 // values and set up a data structure to hold per-BB costs. We cache each 2252 // block's cost so that we don't recompute this when considering different 2253 // subsets of the loop for duplication during unswitching. 2254 SmallPtrSet<const Value *, 4> EphValues; 2255 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues); 2256 SmallDenseMap<BasicBlock *, int, 4> BBCostMap; 2257 2258 // Compute the cost of each block, as well as the total loop cost. Also, bail 2259 // out if we see instructions which are incompatible with loop unswitching 2260 // (convergent, noduplicate, or cross-basic-block tokens). 2261 // FIXME: We might be able to safely handle some of these in non-duplicated 2262 // regions. 2263 int LoopCost = 0; 2264 for (auto *BB : L.blocks()) { 2265 int Cost = 0; 2266 for (auto &I : *BB) { 2267 if (EphValues.count(&I)) 2268 continue; 2269 2270 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB)) 2271 return false; 2272 if (auto CS = CallSite(&I)) 2273 if (CS.isConvergent() || CS.cannotDuplicate()) 2274 return false; 2275 2276 Cost += TTI.getUserCost(&I); 2277 } 2278 assert(Cost >= 0 && "Must not have negative costs!"); 2279 LoopCost += Cost; 2280 assert(LoopCost >= 0 && "Must not have negative loop costs!"); 2281 BBCostMap[BB] = Cost; 2282 } 2283 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n"); 2284 2285 // Now we find the best candidate by searching for the one with the following 2286 // properties in order: 2287 // 2288 // 1) An unswitching cost below the threshold 2289 // 2) The smallest number of duplicated unswitch candidates (to avoid 2290 // creating redundant subsequent unswitching) 2291 // 3) The smallest cost after unswitching. 2292 // 2293 // We prioritize reducing fanout of unswitch candidates provided the cost 2294 // remains below the threshold because this has a multiplicative effect. 2295 // 2296 // This requires memoizing each dominator subtree to avoid redundant work. 2297 // 2298 // FIXME: Need to actually do the number of candidates part above. 2299 SmallDenseMap<DomTreeNode *, int, 4> DTCostMap; 2300 // Given a terminator which might be unswitched, computes the non-duplicated 2301 // cost for that terminator. 2302 auto ComputeUnswitchedCost = [&](TerminatorInst &TI, bool FullUnswitch) { 2303 BasicBlock &BB = *TI.getParent(); 2304 SmallPtrSet<BasicBlock *, 4> Visited; 2305 2306 int Cost = LoopCost; 2307 for (BasicBlock *SuccBB : successors(&BB)) { 2308 // Don't count successors more than once. 2309 if (!Visited.insert(SuccBB).second) 2310 continue; 2311 2312 // If this is a partial unswitch candidate, then it must be a conditional 2313 // branch with a condition of either `or` or `and`. In that case, one of 2314 // the successors is necessarily duplicated, so don't even try to remove 2315 // its cost. 2316 if (!FullUnswitch) { 2317 auto &BI = cast<BranchInst>(TI); 2318 if (cast<Instruction>(BI.getCondition())->getOpcode() == 2319 Instruction::And) { 2320 if (SuccBB == BI.getSuccessor(1)) 2321 continue; 2322 } else { 2323 assert(cast<Instruction>(BI.getCondition())->getOpcode() == 2324 Instruction::Or && 2325 "Only `and` and `or` conditions can result in a partial " 2326 "unswitch!"); 2327 if (SuccBB == BI.getSuccessor(0)) 2328 continue; 2329 } 2330 } 2331 2332 // This successor's domtree will not need to be duplicated after 2333 // unswitching if the edge to the successor dominates it (and thus the 2334 // entire tree). This essentially means there is no other path into this 2335 // subtree and so it will end up live in only one clone of the loop. 2336 if (SuccBB->getUniquePredecessor() || 2337 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 2338 return PredBB == &BB || DT.dominates(SuccBB, PredBB); 2339 })) { 2340 Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap); 2341 assert(Cost >= 0 && 2342 "Non-duplicated cost should never exceed total loop cost!"); 2343 } 2344 } 2345 2346 // Now scale the cost by the number of unique successors minus one. We 2347 // subtract one because there is already at least one copy of the entire 2348 // loop. This is computing the new cost of unswitching a condition. 2349 assert(Visited.size() > 1 && 2350 "Cannot unswitch a condition without multiple distinct successors!"); 2351 return Cost * (Visited.size() - 1); 2352 }; 2353 TerminatorInst *BestUnswitchTI = nullptr; 2354 int BestUnswitchCost; 2355 ArrayRef<Value *> BestUnswitchInvariants; 2356 for (auto &TerminatorAndInvariants : UnswitchCandidates) { 2357 TerminatorInst &TI = *TerminatorAndInvariants.first; 2358 ArrayRef<Value *> Invariants = TerminatorAndInvariants.second; 2359 BranchInst *BI = dyn_cast<BranchInst>(&TI); 2360 int CandidateCost = ComputeUnswitchedCost( 2361 TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 && 2362 Invariants[0] == BI->getCondition())); 2363 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 2364 << " for unswitch candidate: " << TI << "\n"); 2365 if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) { 2366 BestUnswitchTI = &TI; 2367 BestUnswitchCost = CandidateCost; 2368 BestUnswitchInvariants = Invariants; 2369 } 2370 } 2371 2372 if (BestUnswitchCost >= UnswitchThreshold) { 2373 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " 2374 << BestUnswitchCost << "\n"); 2375 return false; 2376 } 2377 2378 LLVM_DEBUG(dbgs() << " Trying to unswitch non-trivial (cost = " 2379 << BestUnswitchCost << ") terminator: " << *BestUnswitchTI 2380 << "\n"); 2381 return unswitchNontrivialInvariants( 2382 L, *BestUnswitchTI, BestUnswitchInvariants, DT, LI, AC, UnswitchCB, SE); 2383 } 2384 2385 /// Unswitch control flow predicated on loop invariant conditions. 2386 /// 2387 /// This first hoists all branches or switches which are trivial (IE, do not 2388 /// require duplicating any part of the loop) out of the loop body. It then 2389 /// looks at other loop invariant control flows and tries to unswitch those as 2390 /// well by cloning the loop if the result is small enough. 2391 /// 2392 /// The `DT`, `LI`, `AC`, `TTI` parameters are required analyses that are also 2393 /// updated based on the unswitch. 2394 /// 2395 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is 2396 /// true, we will attempt to do non-trivial unswitching as well as trivial 2397 /// unswitching. 2398 /// 2399 /// The `UnswitchCB` callback provided will be run after unswitching is 2400 /// complete, with the first parameter set to `true` if the provided loop 2401 /// remains a loop, and a list of new sibling loops created. 2402 /// 2403 /// If `SE` is non-null, we will update that analysis based on the unswitching 2404 /// done. 2405 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, 2406 AssumptionCache &AC, TargetTransformInfo &TTI, 2407 bool NonTrivial, 2408 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB, 2409 ScalarEvolution *SE) { 2410 assert(L.isRecursivelyLCSSAForm(DT, LI) && 2411 "Loops must be in LCSSA form before unswitching."); 2412 bool Changed = false; 2413 2414 // Must be in loop simplified form: we need a preheader and dedicated exits. 2415 if (!L.isLoopSimplifyForm()) 2416 return false; 2417 2418 // Try trivial unswitch first before loop over other basic blocks in the loop. 2419 if (unswitchAllTrivialConditions(L, DT, LI, SE)) { 2420 // If we unswitched successfully we will want to clean up the loop before 2421 // processing it further so just mark it as unswitched and return. 2422 UnswitchCB(/*CurrentLoopValid*/ true, {}); 2423 return true; 2424 } 2425 2426 // If we're not doing non-trivial unswitching, we're done. We both accept 2427 // a parameter but also check a local flag that can be used for testing 2428 // a debugging. 2429 if (!NonTrivial && !EnableNonTrivialUnswitch) 2430 return false; 2431 2432 // For non-trivial unswitching, because it often creates new loops, we rely on 2433 // the pass manager to iterate on the loops rather than trying to immediately 2434 // reach a fixed point. There is no substantial advantage to iterating 2435 // internally, and if any of the new loops are simplified enough to contain 2436 // trivial unswitching we want to prefer those. 2437 2438 // Try to unswitch the best invariant condition. We prefer this full unswitch to 2439 // a partial unswitch when possible below the threshold. 2440 if (unswitchBestCondition(L, DT, LI, AC, TTI, UnswitchCB, SE)) 2441 return true; 2442 2443 // No other opportunities to unswitch. 2444 return Changed; 2445 } 2446 2447 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, 2448 LoopStandardAnalysisResults &AR, 2449 LPMUpdater &U) { 2450 Function &F = *L.getHeader()->getParent(); 2451 (void)F; 2452 2453 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L 2454 << "\n"); 2455 2456 // Save the current loop name in a variable so that we can report it even 2457 // after it has been deleted. 2458 std::string LoopName = L.getName(); 2459 2460 auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid, 2461 ArrayRef<Loop *> NewLoops) { 2462 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2463 if (!NewLoops.empty()) 2464 U.addSiblingLoops(NewLoops); 2465 2466 // If the current loop remains valid, we should revisit it to catch any 2467 // other unswitch opportunities. Otherwise, we need to mark it as deleted. 2468 if (CurrentLoopValid) 2469 U.revisitCurrentLoop(); 2470 else 2471 U.markLoopAsDeleted(L, LoopName); 2472 }; 2473 2474 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, UnswitchCB, 2475 &AR.SE)) 2476 return PreservedAnalyses::all(); 2477 2478 // Historically this pass has had issues with the dominator tree so verify it 2479 // in asserts builds. 2480 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast)); 2481 return getLoopPassPreservedAnalyses(); 2482 } 2483 2484 namespace { 2485 2486 class SimpleLoopUnswitchLegacyPass : public LoopPass { 2487 bool NonTrivial; 2488 2489 public: 2490 static char ID; // Pass ID, replacement for typeid 2491 2492 explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false) 2493 : LoopPass(ID), NonTrivial(NonTrivial) { 2494 initializeSimpleLoopUnswitchLegacyPassPass( 2495 *PassRegistry::getPassRegistry()); 2496 } 2497 2498 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 2499 2500 void getAnalysisUsage(AnalysisUsage &AU) const override { 2501 AU.addRequired<AssumptionCacheTracker>(); 2502 AU.addRequired<TargetTransformInfoWrapperPass>(); 2503 getLoopAnalysisUsage(AU); 2504 } 2505 }; 2506 2507 } // end anonymous namespace 2508 2509 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { 2510 if (skipLoop(L)) 2511 return false; 2512 2513 Function &F = *L->getHeader()->getParent(); 2514 2515 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L 2516 << "\n"); 2517 2518 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2519 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2520 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 2521 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 2522 2523 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); 2524 auto *SE = SEWP ? &SEWP->getSE() : nullptr; 2525 2526 auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid, 2527 ArrayRef<Loop *> NewLoops) { 2528 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2529 for (auto *NewL : NewLoops) 2530 LPM.addLoop(*NewL); 2531 2532 // If the current loop remains valid, re-add it to the queue. This is 2533 // a little wasteful as we'll finish processing the current loop as well, 2534 // but it is the best we can do in the old PM. 2535 if (CurrentLoopValid) 2536 LPM.addLoop(*L); 2537 else 2538 LPM.markLoopAsDeleted(*L); 2539 }; 2540 2541 bool Changed = unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, UnswitchCB, SE); 2542 2543 // If anything was unswitched, also clear any cached information about this 2544 // loop. 2545 LPM.deleteSimpleAnalysisLoop(L); 2546 2547 // Historically this pass has had issues with the dominator tree so verify it 2548 // in asserts builds. 2549 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2550 2551 return Changed; 2552 } 2553 2554 char SimpleLoopUnswitchLegacyPass::ID = 0; 2555 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", 2556 "Simple unswitch loops", false, false) 2557 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 2558 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 2559 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 2560 INITIALIZE_PASS_DEPENDENCY(LoopPass) 2561 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 2562 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", 2563 "Simple unswitch loops", false, false) 2564 2565 Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) { 2566 return new SimpleLoopUnswitchLegacyPass(NonTrivial); 2567 } 2568