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