1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 10 #include "llvm/ADT/DenseMap.h" 11 #include "llvm/ADT/STLExtras.h" 12 #include "llvm/ADT/Sequence.h" 13 #include "llvm/ADT/SetVector.h" 14 #include "llvm/ADT/SmallPtrSet.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/Statistic.h" 17 #include "llvm/ADT/Twine.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/CFG.h" 20 #include "llvm/Analysis/CodeMetrics.h" 21 #include "llvm/Analysis/GuardUtils.h" 22 #include "llvm/Analysis/InstructionSimplify.h" 23 #include "llvm/Analysis/LoopAnalysisManager.h" 24 #include "llvm/Analysis/LoopInfo.h" 25 #include "llvm/Analysis/LoopIterator.h" 26 #include "llvm/Analysis/LoopPass.h" 27 #include "llvm/Analysis/MemorySSA.h" 28 #include "llvm/Analysis/MemorySSAUpdater.h" 29 #include "llvm/Analysis/Utils/Local.h" 30 #include "llvm/IR/BasicBlock.h" 31 #include "llvm/IR/Constant.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/IR/Dominators.h" 34 #include "llvm/IR/Function.h" 35 #include "llvm/IR/InstrTypes.h" 36 #include "llvm/IR/Instruction.h" 37 #include "llvm/IR/Instructions.h" 38 #include "llvm/IR/IntrinsicInst.h" 39 #include "llvm/IR/Use.h" 40 #include "llvm/IR/Value.h" 41 #include "llvm/InitializePasses.h" 42 #include "llvm/Pass.h" 43 #include "llvm/Support/Casting.h" 44 #include "llvm/Support/Debug.h" 45 #include "llvm/Support/ErrorHandling.h" 46 #include "llvm/Support/GenericDomTree.h" 47 #include "llvm/Support/raw_ostream.h" 48 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 49 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 50 #include "llvm/Transforms/Utils/Cloning.h" 51 #include "llvm/Transforms/Utils/LoopUtils.h" 52 #include "llvm/Transforms/Utils/ValueMapper.h" 53 #include <algorithm> 54 #include <cassert> 55 #include <iterator> 56 #include <numeric> 57 #include <utility> 58 59 #define DEBUG_TYPE "simple-loop-unswitch" 60 61 using namespace llvm; 62 63 STATISTIC(NumBranches, "Number of branches unswitched"); 64 STATISTIC(NumSwitches, "Number of switches unswitched"); 65 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching"); 66 STATISTIC(NumTrivial, "Number of unswitches that are trivial"); 67 STATISTIC( 68 NumCostMultiplierSkipped, 69 "Number of unswitch candidates that had their cost multiplier skipped"); 70 71 static cl::opt<bool> EnableNonTrivialUnswitch( 72 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden, 73 cl::desc("Forcibly enables non-trivial loop unswitching rather than " 74 "following the configuration passed into the pass.")); 75 76 static cl::opt<int> 77 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, 78 cl::desc("The cost threshold for unswitching a loop.")); 79 80 static cl::opt<bool> EnableUnswitchCostMultiplier( 81 "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden, 82 cl::desc("Enable unswitch cost multiplier that prohibits exponential " 83 "explosion in nontrivial unswitch.")); 84 static cl::opt<int> UnswitchSiblingsToplevelDiv( 85 "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden, 86 cl::desc("Toplevel siblings divisor for cost multiplier.")); 87 static cl::opt<int> UnswitchNumInitialUnscaledCandidates( 88 "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden, 89 cl::desc("Number of unswitch candidates that are ignored when calculating " 90 "cost multiplier.")); 91 static cl::opt<bool> UnswitchGuards( 92 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden, 93 cl::desc("If enabled, simple loop unswitching will also consider " 94 "llvm.experimental.guard intrinsics as unswitch candidates.")); 95 96 /// Collect all of the loop invariant input values transitively used by the 97 /// homogeneous instruction graph from a given root. 98 /// 99 /// This essentially walks from a root recursively through loop variant operands 100 /// which have the exact same opcode and finds all inputs which are loop 101 /// invariant. For some operations these can be re-associated and unswitched out 102 /// of the loop entirely. 103 static TinyPtrVector<Value *> 104 collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root, 105 LoopInfo &LI) { 106 assert(!L.isLoopInvariant(&Root) && 107 "Only need to walk the graph if root itself is not invariant."); 108 TinyPtrVector<Value *> Invariants; 109 110 // Build a worklist and recurse through operators collecting invariants. 111 SmallVector<Instruction *, 4> Worklist; 112 SmallPtrSet<Instruction *, 8> Visited; 113 Worklist.push_back(&Root); 114 Visited.insert(&Root); 115 do { 116 Instruction &I = *Worklist.pop_back_val(); 117 for (Value *OpV : I.operand_values()) { 118 // Skip constants as unswitching isn't interesting for them. 119 if (isa<Constant>(OpV)) 120 continue; 121 122 // Add it to our result if loop invariant. 123 if (L.isLoopInvariant(OpV)) { 124 Invariants.push_back(OpV); 125 continue; 126 } 127 128 // If not an instruction with the same opcode, nothing we can do. 129 Instruction *OpI = dyn_cast<Instruction>(OpV); 130 if (!OpI || OpI->getOpcode() != Root.getOpcode()) 131 continue; 132 133 // Visit this operand. 134 if (Visited.insert(OpI).second) 135 Worklist.push_back(OpI); 136 } 137 } while (!Worklist.empty()); 138 139 return Invariants; 140 } 141 142 static void replaceLoopInvariantUses(Loop &L, Value *Invariant, 143 Constant &Replacement) { 144 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?"); 145 146 // Replace uses of LIC in the loop with the given constant. 147 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); UI != UE;) { 148 // Grab the use and walk past it so we can clobber it in the use list. 149 Use *U = &*UI++; 150 Instruction *UserI = dyn_cast<Instruction>(U->getUser()); 151 152 // Replace this use within the loop body. 153 if (UserI && L.contains(UserI)) 154 U->set(&Replacement); 155 } 156 } 157 158 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial 159 /// incoming values along this edge. 160 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB, 161 BasicBlock &ExitBB) { 162 for (Instruction &I : ExitBB) { 163 auto *PN = dyn_cast<PHINode>(&I); 164 if (!PN) 165 // No more PHIs to check. 166 return true; 167 168 // If the incoming value for this edge isn't loop invariant the unswitch 169 // won't be trivial. 170 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))) 171 return false; 172 } 173 llvm_unreachable("Basic blocks should never be empty!"); 174 } 175 176 /// Insert code to test a set of loop invariant values, and conditionally branch 177 /// on them. 178 static void buildPartialUnswitchConditionalBranch(BasicBlock &BB, 179 ArrayRef<Value *> Invariants, 180 bool Direction, 181 BasicBlock &UnswitchedSucc, 182 BasicBlock &NormalSucc) { 183 IRBuilder<> IRB(&BB); 184 185 Value *Cond = Direction ? IRB.CreateOr(Invariants) : 186 IRB.CreateAnd(Invariants); 187 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, 188 Direction ? &NormalSucc : &UnswitchedSucc); 189 } 190 191 /// Rewrite the PHI nodes in an unswitched loop exit basic block. 192 /// 193 /// Requires that the loop exit and unswitched basic block are the same, and 194 /// that the exiting block was a unique predecessor of that block. Rewrites the 195 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial 196 /// PHI nodes from the old preheader that now contains the unswitched 197 /// terminator. 198 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, 199 BasicBlock &OldExitingBB, 200 BasicBlock &OldPH) { 201 for (PHINode &PN : UnswitchedBB.phis()) { 202 // When the loop exit is directly unswitched we just need to update the 203 // incoming basic block. We loop to handle weird cases with repeated 204 // incoming blocks, but expect to typically only have one operand here. 205 for (auto i : seq<int>(0, PN.getNumOperands())) { 206 assert(PN.getIncomingBlock(i) == &OldExitingBB && 207 "Found incoming block different from unique predecessor!"); 208 PN.setIncomingBlock(i, &OldPH); 209 } 210 } 211 } 212 213 /// Rewrite the PHI nodes in the loop exit basic block and the split off 214 /// unswitched block. 215 /// 216 /// Because the exit block remains an exit from the loop, this rewrites the 217 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI 218 /// nodes into the unswitched basic block to select between the value in the 219 /// old preheader and the loop exit. 220 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, 221 BasicBlock &UnswitchedBB, 222 BasicBlock &OldExitingBB, 223 BasicBlock &OldPH, 224 bool FullUnswitch) { 225 assert(&ExitBB != &UnswitchedBB && 226 "Must have different loop exit and unswitched blocks!"); 227 Instruction *InsertPt = &*UnswitchedBB.begin(); 228 for (PHINode &PN : ExitBB.phis()) { 229 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2, 230 PN.getName() + ".split", InsertPt); 231 232 // Walk backwards over the old PHI node's inputs to minimize the cost of 233 // removing each one. We have to do this weird loop manually so that we 234 // create the same number of new incoming edges in the new PHI as we expect 235 // each case-based edge to be included in the unswitched switch in some 236 // cases. 237 // FIXME: This is really, really gross. It would be much cleaner if LLVM 238 // allowed us to create a single entry for a predecessor block without 239 // having separate entries for each "edge" even though these edges are 240 // required to produce identical results. 241 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) { 242 if (PN.getIncomingBlock(i) != &OldExitingBB) 243 continue; 244 245 Value *Incoming = PN.getIncomingValue(i); 246 if (FullUnswitch) 247 // No more edge from the old exiting block to the exit block. 248 PN.removeIncomingValue(i); 249 250 NewPN->addIncoming(Incoming, &OldPH); 251 } 252 253 // Now replace the old PHI with the new one and wire the old one in as an 254 // input to the new one. 255 PN.replaceAllUsesWith(NewPN); 256 NewPN->addIncoming(&PN, &ExitBB); 257 } 258 } 259 260 /// Hoist the current loop up to the innermost loop containing a remaining exit. 261 /// 262 /// Because we've removed an exit from the loop, we may have changed the set of 263 /// loops reachable and need to move the current loop up the loop nest or even 264 /// to an entirely separate nest. 265 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader, 266 DominatorTree &DT, LoopInfo &LI, 267 MemorySSAUpdater *MSSAU) { 268 // If the loop is already at the top level, we can't hoist it anywhere. 269 Loop *OldParentL = L.getParentLoop(); 270 if (!OldParentL) 271 return; 272 273 SmallVector<BasicBlock *, 4> Exits; 274 L.getExitBlocks(Exits); 275 Loop *NewParentL = nullptr; 276 for (auto *ExitBB : Exits) 277 if (Loop *ExitL = LI.getLoopFor(ExitBB)) 278 if (!NewParentL || NewParentL->contains(ExitL)) 279 NewParentL = ExitL; 280 281 if (NewParentL == OldParentL) 282 return; 283 284 // The new parent loop (if different) should always contain the old one. 285 if (NewParentL) 286 assert(NewParentL->contains(OldParentL) && 287 "Can only hoist this loop up the nest!"); 288 289 // The preheader will need to move with the body of this loop. However, 290 // because it isn't in this loop we also need to update the primary loop map. 291 assert(OldParentL == LI.getLoopFor(&Preheader) && 292 "Parent loop of this loop should contain this loop's preheader!"); 293 LI.changeLoopFor(&Preheader, NewParentL); 294 295 // Remove this loop from its old parent. 296 OldParentL->removeChildLoop(&L); 297 298 // Add the loop either to the new parent or as a top-level loop. 299 if (NewParentL) 300 NewParentL->addChildLoop(&L); 301 else 302 LI.addTopLevelLoop(&L); 303 304 // Remove this loops blocks from the old parent and every other loop up the 305 // nest until reaching the new parent. Also update all of these 306 // no-longer-containing loops to reflect the nesting change. 307 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL; 308 OldContainingL = OldContainingL->getParentLoop()) { 309 llvm::erase_if(OldContainingL->getBlocksVector(), 310 [&](const BasicBlock *BB) { 311 return BB == &Preheader || L.contains(BB); 312 }); 313 314 OldContainingL->getBlocksSet().erase(&Preheader); 315 for (BasicBlock *BB : L.blocks()) 316 OldContainingL->getBlocksSet().erase(BB); 317 318 // Because we just hoisted a loop out of this one, we have essentially 319 // created new exit paths from it. That means we need to form LCSSA PHI 320 // nodes for values used in the no-longer-nested loop. 321 formLCSSA(*OldContainingL, DT, &LI, nullptr); 322 323 // We shouldn't need to form dedicated exits because the exit introduced 324 // here is the (just split by unswitching) preheader. However, after trivial 325 // unswitching it is possible to get new non-dedicated exits out of parent 326 // loop so let's conservatively form dedicated exit blocks and figure out 327 // if we can optimize later. 328 formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU, 329 /*PreserveLCSSA*/ true); 330 } 331 } 332 333 /// Unswitch a trivial branch if the condition is loop invariant. 334 /// 335 /// This routine should only be called when loop code leading to the branch has 336 /// been validated as trivial (no side effects). This routine checks if the 337 /// condition is invariant and one of the successors is a loop exit. This 338 /// allows us to unswitch without duplicating the loop, making it trivial. 339 /// 340 /// If this routine fails to unswitch the branch it returns false. 341 /// 342 /// If the branch can be unswitched, this routine splits the preheader and 343 /// hoists the branch above that split. Preserves loop simplified form 344 /// (splitting the exit block as necessary). It simplifies the branch within 345 /// the loop to an unconditional branch but doesn't remove it entirely. Further 346 /// cleanup can be done with some simplify-cfg like pass. 347 /// 348 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 349 /// invalidated by this. 350 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, 351 LoopInfo &LI, ScalarEvolution *SE, 352 MemorySSAUpdater *MSSAU) { 353 assert(BI.isConditional() && "Can only unswitch a conditional branch!"); 354 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n"); 355 356 // The loop invariant values that we want to unswitch. 357 TinyPtrVector<Value *> Invariants; 358 359 // When true, we're fully unswitching the branch rather than just unswitching 360 // some input conditions to the branch. 361 bool FullUnswitch = false; 362 363 if (L.isLoopInvariant(BI.getCondition())) { 364 Invariants.push_back(BI.getCondition()); 365 FullUnswitch = true; 366 } else { 367 if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition())) 368 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI); 369 if (Invariants.empty()) 370 // Couldn't find invariant inputs! 371 return false; 372 } 373 374 // Check that one of the branch's successors exits, and which one. 375 bool ExitDirection = true; 376 int LoopExitSuccIdx = 0; 377 auto *LoopExitBB = BI.getSuccessor(0); 378 if (L.contains(LoopExitBB)) { 379 ExitDirection = false; 380 LoopExitSuccIdx = 1; 381 LoopExitBB = BI.getSuccessor(1); 382 if (L.contains(LoopExitBB)) 383 return false; 384 } 385 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); 386 auto *ParentBB = BI.getParent(); 387 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) 388 return false; 389 390 // When unswitching only part of the branch's condition, we need the exit 391 // block to be reached directly from the partially unswitched input. This can 392 // be done when the exit block is along the true edge and the branch condition 393 // is a graph of `or` operations, or the exit block is along the false edge 394 // and the condition is a graph of `and` operations. 395 if (!FullUnswitch) { 396 if (ExitDirection) { 397 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::Or) 398 return false; 399 } else { 400 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::And) 401 return false; 402 } 403 } 404 405 LLVM_DEBUG({ 406 dbgs() << " unswitching trivial invariant conditions for: " << BI 407 << "\n"; 408 for (Value *Invariant : Invariants) { 409 dbgs() << " " << *Invariant << " == true"; 410 if (Invariant != Invariants.back()) 411 dbgs() << " ||"; 412 dbgs() << "\n"; 413 } 414 }); 415 416 // If we have scalar evolutions, we need to invalidate them including this 417 // loop and the loop containing the exit block. 418 if (SE) { 419 if (Loop *ExitL = LI.getLoopFor(LoopExitBB)) 420 SE->forgetLoop(ExitL); 421 else 422 // Forget the entire nest as this exits the entire nest. 423 SE->forgetTopmostLoop(&L); 424 } 425 426 if (MSSAU && VerifyMemorySSA) 427 MSSAU->getMemorySSA()->verifyMemorySSA(); 428 429 // Split the preheader, so that we know that there is a safe place to insert 430 // the conditional branch. We will change the preheader to have a conditional 431 // branch on LoopCond. 432 BasicBlock *OldPH = L.getLoopPreheader(); 433 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); 434 435 // Now that we have a place to insert the conditional branch, create a place 436 // to branch to: this is the exit block out of the loop that we are 437 // unswitching. We need to split this if there are other loop predecessors. 438 // Because the loop is in simplified form, *any* other predecessor is enough. 439 BasicBlock *UnswitchedBB; 440 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) { 441 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() && 442 "A branch's parent isn't a predecessor!"); 443 UnswitchedBB = LoopExitBB; 444 } else { 445 UnswitchedBB = 446 SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI, MSSAU); 447 } 448 449 if (MSSAU && VerifyMemorySSA) 450 MSSAU->getMemorySSA()->verifyMemorySSA(); 451 452 // Actually move the invariant uses into the unswitched position. If possible, 453 // we do this by moving the instructions, but when doing partial unswitching 454 // we do it by building a new merge of the values in the unswitched position. 455 OldPH->getTerminator()->eraseFromParent(); 456 if (FullUnswitch) { 457 // If fully unswitching, we can use the existing branch instruction. 458 // Splice it into the old PH to gate reaching the new preheader and re-point 459 // its successors. 460 OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(), 461 BI); 462 if (MSSAU) { 463 // Temporarily clone the terminator, to make MSSA update cheaper by 464 // separating "insert edge" updates from "remove edge" ones. 465 ParentBB->getInstList().push_back(BI.clone()); 466 } else { 467 // Create a new unconditional branch that will continue the loop as a new 468 // terminator. 469 BranchInst::Create(ContinueBB, ParentBB); 470 } 471 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); 472 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); 473 } else { 474 // Only unswitching a subset of inputs to the condition, so we will need to 475 // build a new branch that merges the invariant inputs. 476 if (ExitDirection) 477 assert(cast<Instruction>(BI.getCondition())->getOpcode() == 478 Instruction::Or && 479 "Must have an `or` of `i1`s for the condition!"); 480 else 481 assert(cast<Instruction>(BI.getCondition())->getOpcode() == 482 Instruction::And && 483 "Must have an `and` of `i1`s for the condition!"); 484 buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection, 485 *UnswitchedBB, *NewPH); 486 } 487 488 // Update the dominator tree with the added edge. 489 DT.insertEdge(OldPH, UnswitchedBB); 490 491 // After the dominator tree was updated with the added edge, update MemorySSA 492 // if available. 493 if (MSSAU) { 494 SmallVector<CFGUpdate, 1> Updates; 495 Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB}); 496 MSSAU->applyInsertUpdates(Updates, DT); 497 } 498 499 // Finish updating dominator tree and memory ssa for full unswitch. 500 if (FullUnswitch) { 501 if (MSSAU) { 502 // Remove the cloned branch instruction. 503 ParentBB->getTerminator()->eraseFromParent(); 504 // Create unconditional branch now. 505 BranchInst::Create(ContinueBB, ParentBB); 506 MSSAU->removeEdge(ParentBB, LoopExitBB); 507 } 508 DT.deleteEdge(ParentBB, LoopExitBB); 509 } 510 511 if (MSSAU && VerifyMemorySSA) 512 MSSAU->getMemorySSA()->verifyMemorySSA(); 513 514 // Rewrite the relevant PHI nodes. 515 if (UnswitchedBB == LoopExitBB) 516 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); 517 else 518 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, 519 *ParentBB, *OldPH, FullUnswitch); 520 521 // The constant we can replace all of our invariants with inside the loop 522 // body. If any of the invariants have a value other than this the loop won't 523 // be entered. 524 ConstantInt *Replacement = ExitDirection 525 ? ConstantInt::getFalse(BI.getContext()) 526 : ConstantInt::getTrue(BI.getContext()); 527 528 // Since this is an i1 condition we can also trivially replace uses of it 529 // within the loop with a constant. 530 for (Value *Invariant : Invariants) 531 replaceLoopInvariantUses(L, Invariant, *Replacement); 532 533 // If this was full unswitching, we may have changed the nesting relationship 534 // for this loop so hoist it to its correct parent if needed. 535 if (FullUnswitch) 536 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU); 537 538 if (MSSAU && VerifyMemorySSA) 539 MSSAU->getMemorySSA()->verifyMemorySSA(); 540 541 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n"); 542 ++NumTrivial; 543 ++NumBranches; 544 return true; 545 } 546 547 /// Unswitch a trivial switch if the condition is loop invariant. 548 /// 549 /// This routine should only be called when loop code leading to the switch has 550 /// been validated as trivial (no side effects). This routine checks if the 551 /// condition is invariant and that at least one of the successors is a loop 552 /// exit. This allows us to unswitch without duplicating the loop, making it 553 /// trivial. 554 /// 555 /// If this routine fails to unswitch the switch it returns false. 556 /// 557 /// If the switch can be unswitched, this routine splits the preheader and 558 /// copies the switch above that split. If the default case is one of the 559 /// exiting cases, it copies the non-exiting cases and points them at the new 560 /// preheader. If the default case is not exiting, it copies the exiting cases 561 /// and points the default at the preheader. It preserves loop simplified form 562 /// (splitting the exit blocks as necessary). It simplifies the switch within 563 /// the loop by removing now-dead cases. If the default case is one of those 564 /// unswitched, it replaces its destination with a new basic block containing 565 /// only unreachable. Such basic blocks, while technically loop exits, are not 566 /// considered for unswitching so this is a stable transform and the same 567 /// switch will not be revisited. If after unswitching there is only a single 568 /// in-loop successor, the switch is further simplified to an unconditional 569 /// branch. Still more cleanup can be done with some simplify-cfg like pass. 570 /// 571 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 572 /// invalidated by this. 573 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, 574 LoopInfo &LI, ScalarEvolution *SE, 575 MemorySSAUpdater *MSSAU) { 576 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n"); 577 Value *LoopCond = SI.getCondition(); 578 579 // If this isn't switching on an invariant condition, we can't unswitch it. 580 if (!L.isLoopInvariant(LoopCond)) 581 return false; 582 583 auto *ParentBB = SI.getParent(); 584 585 SmallVector<int, 4> ExitCaseIndices; 586 for (auto Case : SI.cases()) { 587 auto *SuccBB = Case.getCaseSuccessor(); 588 if (!L.contains(SuccBB) && 589 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB)) 590 ExitCaseIndices.push_back(Case.getCaseIndex()); 591 } 592 BasicBlock *DefaultExitBB = nullptr; 593 SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight = 594 SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0); 595 if (!L.contains(SI.getDefaultDest()) && 596 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) && 597 !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator())) { 598 DefaultExitBB = SI.getDefaultDest(); 599 } else if (ExitCaseIndices.empty()) 600 return false; 601 602 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n"); 603 604 if (MSSAU && VerifyMemorySSA) 605 MSSAU->getMemorySSA()->verifyMemorySSA(); 606 607 // We may need to invalidate SCEVs for the outermost loop reached by any of 608 // the exits. 609 Loop *OuterL = &L; 610 611 if (DefaultExitBB) { 612 // Clear out the default destination temporarily to allow accurate 613 // predecessor lists to be examined below. 614 SI.setDefaultDest(nullptr); 615 // Check the loop containing this exit. 616 Loop *ExitL = LI.getLoopFor(DefaultExitBB); 617 if (!ExitL || ExitL->contains(OuterL)) 618 OuterL = ExitL; 619 } 620 621 // Store the exit cases into a separate data structure and remove them from 622 // the switch. 623 SmallVector<std::tuple<ConstantInt *, BasicBlock *, 624 SwitchInstProfUpdateWrapper::CaseWeightOpt>, 625 4> ExitCases; 626 ExitCases.reserve(ExitCaseIndices.size()); 627 SwitchInstProfUpdateWrapper SIW(SI); 628 // We walk the case indices backwards so that we remove the last case first 629 // and don't disrupt the earlier indices. 630 for (unsigned Index : reverse(ExitCaseIndices)) { 631 auto CaseI = SI.case_begin() + Index; 632 // Compute the outer loop from this exit. 633 Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor()); 634 if (!ExitL || ExitL->contains(OuterL)) 635 OuterL = ExitL; 636 // Save the value of this case. 637 auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex()); 638 ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W); 639 // Delete the unswitched cases. 640 SIW.removeCase(CaseI); 641 } 642 643 if (SE) { 644 if (OuterL) 645 SE->forgetLoop(OuterL); 646 else 647 SE->forgetTopmostLoop(&L); 648 } 649 650 // Check if after this all of the remaining cases point at the same 651 // successor. 652 BasicBlock *CommonSuccBB = nullptr; 653 if (SI.getNumCases() > 0 && 654 std::all_of(std::next(SI.case_begin()), SI.case_end(), 655 [&SI](const SwitchInst::CaseHandle &Case) { 656 return Case.getCaseSuccessor() == 657 SI.case_begin()->getCaseSuccessor(); 658 })) 659 CommonSuccBB = SI.case_begin()->getCaseSuccessor(); 660 if (!DefaultExitBB) { 661 // If we're not unswitching the default, we need it to match any cases to 662 // have a common successor or if we have no cases it is the common 663 // successor. 664 if (SI.getNumCases() == 0) 665 CommonSuccBB = SI.getDefaultDest(); 666 else if (SI.getDefaultDest() != CommonSuccBB) 667 CommonSuccBB = nullptr; 668 } 669 670 // Split the preheader, so that we know that there is a safe place to insert 671 // the switch. 672 BasicBlock *OldPH = L.getLoopPreheader(); 673 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); 674 OldPH->getTerminator()->eraseFromParent(); 675 676 // Now add the unswitched switch. 677 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); 678 SwitchInstProfUpdateWrapper NewSIW(*NewSI); 679 680 // Rewrite the IR for the unswitched basic blocks. This requires two steps. 681 // First, we split any exit blocks with remaining in-loop predecessors. Then 682 // we update the PHIs in one of two ways depending on if there was a split. 683 // We walk in reverse so that we split in the same order as the cases 684 // appeared. This is purely for convenience of reading the resulting IR, but 685 // it doesn't cost anything really. 686 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; 687 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; 688 // Handle the default exit if necessary. 689 // FIXME: It'd be great if we could merge this with the loop below but LLVM's 690 // ranges aren't quite powerful enough yet. 691 if (DefaultExitBB) { 692 if (pred_empty(DefaultExitBB)) { 693 UnswitchedExitBBs.insert(DefaultExitBB); 694 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); 695 } else { 696 auto *SplitBB = 697 SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI, MSSAU); 698 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB, 699 *ParentBB, *OldPH, 700 /*FullUnswitch*/ true); 701 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; 702 } 703 } 704 // Note that we must use a reference in the for loop so that we update the 705 // container. 706 for (auto &ExitCase : reverse(ExitCases)) { 707 // Grab a reference to the exit block in the pair so that we can update it. 708 BasicBlock *ExitBB = std::get<1>(ExitCase); 709 710 // If this case is the last edge into the exit block, we can simply reuse it 711 // as it will no longer be a loop exit. No mapping necessary. 712 if (pred_empty(ExitBB)) { 713 // Only rewrite once. 714 if (UnswitchedExitBBs.insert(ExitBB).second) 715 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); 716 continue; 717 } 718 719 // Otherwise we need to split the exit block so that we retain an exit 720 // block from the loop and a target for the unswitched condition. 721 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; 722 if (!SplitExitBB) { 723 // If this is the first time we see this, do the split and remember it. 724 SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU); 725 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB, 726 *ParentBB, *OldPH, 727 /*FullUnswitch*/ true); 728 } 729 // Update the case pair to point to the split block. 730 std::get<1>(ExitCase) = SplitExitBB; 731 } 732 733 // Now add the unswitched cases. We do this in reverse order as we built them 734 // in reverse order. 735 for (auto &ExitCase : reverse(ExitCases)) { 736 ConstantInt *CaseVal = std::get<0>(ExitCase); 737 BasicBlock *UnswitchedBB = std::get<1>(ExitCase); 738 739 NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase)); 740 } 741 742 // If the default was unswitched, re-point it and add explicit cases for 743 // entering the loop. 744 if (DefaultExitBB) { 745 NewSIW->setDefaultDest(DefaultExitBB); 746 NewSIW.setSuccessorWeight(0, DefaultCaseWeight); 747 748 // We removed all the exit cases, so we just copy the cases to the 749 // unswitched switch. 750 for (const auto &Case : SI.cases()) 751 NewSIW.addCase(Case.getCaseValue(), NewPH, 752 SIW.getSuccessorWeight(Case.getSuccessorIndex())); 753 } else if (DefaultCaseWeight) { 754 // We have to set branch weight of the default case. 755 uint64_t SW = *DefaultCaseWeight; 756 for (const auto &Case : SI.cases()) { 757 auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex()); 758 assert(W && 759 "case weight must be defined as default case weight is defined"); 760 SW += *W; 761 } 762 NewSIW.setSuccessorWeight(0, SW); 763 } 764 765 // If we ended up with a common successor for every path through the switch 766 // after unswitching, rewrite it to an unconditional branch to make it easy 767 // to recognize. Otherwise we potentially have to recognize the default case 768 // pointing at unreachable and other complexity. 769 if (CommonSuccBB) { 770 BasicBlock *BB = SI.getParent(); 771 // We may have had multiple edges to this common successor block, so remove 772 // them as predecessors. We skip the first one, either the default or the 773 // actual first case. 774 bool SkippedFirst = DefaultExitBB == nullptr; 775 for (auto Case : SI.cases()) { 776 assert(Case.getCaseSuccessor() == CommonSuccBB && 777 "Non-common successor!"); 778 (void)Case; 779 if (!SkippedFirst) { 780 SkippedFirst = true; 781 continue; 782 } 783 CommonSuccBB->removePredecessor(BB, 784 /*KeepOneInputPHIs*/ true); 785 } 786 // Now nuke the switch and replace it with a direct branch. 787 SIW.eraseFromParent(); 788 BranchInst::Create(CommonSuccBB, BB); 789 } else if (DefaultExitBB) { 790 assert(SI.getNumCases() > 0 && 791 "If we had no cases we'd have a common successor!"); 792 // Move the last case to the default successor. This is valid as if the 793 // default got unswitched it cannot be reached. This has the advantage of 794 // being simple and keeping the number of edges from this switch to 795 // successors the same, and avoiding any PHI update complexity. 796 auto LastCaseI = std::prev(SI.case_end()); 797 798 SI.setDefaultDest(LastCaseI->getCaseSuccessor()); 799 SIW.setSuccessorWeight( 800 0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex())); 801 SIW.removeCase(LastCaseI); 802 } 803 804 // Walk the unswitched exit blocks and the unswitched split blocks and update 805 // the dominator tree based on the CFG edits. While we are walking unordered 806 // containers here, the API for applyUpdates takes an unordered list of 807 // updates and requires them to not contain duplicates. 808 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 809 for (auto *UnswitchedExitBB : UnswitchedExitBBs) { 810 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB}); 811 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB}); 812 } 813 for (auto SplitUnswitchedPair : SplitExitBBMap) { 814 DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first}); 815 DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second}); 816 } 817 DT.applyUpdates(DTUpdates); 818 819 if (MSSAU) { 820 MSSAU->applyUpdates(DTUpdates, DT); 821 if (VerifyMemorySSA) 822 MSSAU->getMemorySSA()->verifyMemorySSA(); 823 } 824 825 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 826 827 // We may have changed the nesting relationship for this loop so hoist it to 828 // its correct parent if needed. 829 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU); 830 831 if (MSSAU && VerifyMemorySSA) 832 MSSAU->getMemorySSA()->verifyMemorySSA(); 833 834 ++NumTrivial; 835 ++NumSwitches; 836 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n"); 837 return true; 838 } 839 840 /// This routine scans the loop to find a branch or switch which occurs before 841 /// any side effects occur. These can potentially be unswitched without 842 /// duplicating the loop. If a branch or switch is successfully unswitched the 843 /// scanning continues to see if subsequent branches or switches have become 844 /// trivial. Once all trivial candidates have been unswitched, this routine 845 /// returns. 846 /// 847 /// The return value indicates whether anything was unswitched (and therefore 848 /// changed). 849 /// 850 /// If `SE` is not null, it will be updated based on the potential loop SCEVs 851 /// invalidated by this. 852 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, 853 LoopInfo &LI, ScalarEvolution *SE, 854 MemorySSAUpdater *MSSAU) { 855 bool Changed = false; 856 857 // If loop header has only one reachable successor we should keep looking for 858 // trivial condition candidates in the successor as well. An alternative is 859 // to constant fold conditions and merge successors into loop header (then we 860 // only need to check header's terminator). The reason for not doing this in 861 // LoopUnswitch pass is that it could potentially break LoopPassManager's 862 // invariants. Folding dead branches could either eliminate the current loop 863 // or make other loops unreachable. LCSSA form might also not be preserved 864 // after deleting branches. The following code keeps traversing loop header's 865 // successors until it finds the trivial condition candidate (condition that 866 // is not a constant). Since unswitching generates branches with constant 867 // conditions, this scenario could be very common in practice. 868 BasicBlock *CurrentBB = L.getHeader(); 869 SmallPtrSet<BasicBlock *, 8> Visited; 870 Visited.insert(CurrentBB); 871 do { 872 // Check if there are any side-effecting instructions (e.g. stores, calls, 873 // volatile loads) in the part of the loop that the code *would* execute 874 // without unswitching. 875 if (MSSAU) // Possible early exit with MSSA 876 if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB)) 877 if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end())) 878 return Changed; 879 if (llvm::any_of(*CurrentBB, 880 [](Instruction &I) { return I.mayHaveSideEffects(); })) 881 return Changed; 882 883 Instruction *CurrentTerm = CurrentBB->getTerminator(); 884 885 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) { 886 // Don't bother trying to unswitch past a switch with a constant 887 // condition. This should be removed prior to running this pass by 888 // simplify-cfg. 889 if (isa<Constant>(SI->getCondition())) 890 return Changed; 891 892 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU)) 893 // Couldn't unswitch this one so we're done. 894 return Changed; 895 896 // Mark that we managed to unswitch something. 897 Changed = true; 898 899 // If unswitching turned the terminator into an unconditional branch then 900 // we can continue. The unswitching logic specifically works to fold any 901 // cases it can into an unconditional branch to make it easier to 902 // recognize here. 903 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); 904 if (!BI || BI->isConditional()) 905 return Changed; 906 907 CurrentBB = BI->getSuccessor(0); 908 continue; 909 } 910 911 auto *BI = dyn_cast<BranchInst>(CurrentTerm); 912 if (!BI) 913 // We do not understand other terminator instructions. 914 return Changed; 915 916 // Don't bother trying to unswitch past an unconditional branch or a branch 917 // with a constant value. These should be removed by simplify-cfg prior to 918 // running this pass. 919 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 920 return Changed; 921 922 // Found a trivial condition candidate: non-foldable conditional branch. If 923 // we fail to unswitch this, we can't do anything else that is trivial. 924 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU)) 925 return Changed; 926 927 // Mark that we managed to unswitch something. 928 Changed = true; 929 930 // If we only unswitched some of the conditions feeding the branch, we won't 931 // have collapsed it to a single successor. 932 BI = cast<BranchInst>(CurrentBB->getTerminator()); 933 if (BI->isConditional()) 934 return Changed; 935 936 // Follow the newly unconditional branch into its successor. 937 CurrentBB = BI->getSuccessor(0); 938 939 // When continuing, if we exit the loop or reach a previous visited block, 940 // then we can not reach any trivial condition candidates (unfoldable 941 // branch instructions or switch instructions) and no unswitch can happen. 942 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second); 943 944 return Changed; 945 } 946 947 /// Build the cloned blocks for an unswitched copy of the given loop. 948 /// 949 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and 950 /// after the split block (`SplitBB`) that will be used to select between the 951 /// cloned and original loop. 952 /// 953 /// This routine handles cloning all of the necessary loop blocks and exit 954 /// blocks including rewriting their instructions and the relevant PHI nodes. 955 /// Any loop blocks or exit blocks which are dominated by a different successor 956 /// than the one for this clone of the loop blocks can be trivially skipped. We 957 /// use the `DominatingSucc` map to determine whether a block satisfies that 958 /// property with a simple map lookup. 959 /// 960 /// It also correctly creates the unconditional branch in the cloned 961 /// unswitched parent block to only point at the unswitched successor. 962 /// 963 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit 964 /// block splitting is correctly reflected in `LoopInfo`, essentially all of 965 /// the cloned blocks (and their loops) are left without full `LoopInfo` 966 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned 967 /// blocks to them but doesn't create the cloned `DominatorTree` structure and 968 /// instead the caller must recompute an accurate DT. It *does* correctly 969 /// update the `AssumptionCache` provided in `AC`. 970 static BasicBlock *buildClonedLoopBlocks( 971 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, 972 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB, 973 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, 974 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc, 975 ValueToValueMapTy &VMap, 976 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC, 977 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) { 978 SmallVector<BasicBlock *, 4> NewBlocks; 979 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size()); 980 981 // We will need to clone a bunch of blocks, wrap up the clone operation in 982 // a helper. 983 auto CloneBlock = [&](BasicBlock *OldBB) { 984 // Clone the basic block and insert it before the new preheader. 985 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent()); 986 NewBB->moveBefore(LoopPH); 987 988 // Record this block and the mapping. 989 NewBlocks.push_back(NewBB); 990 VMap[OldBB] = NewBB; 991 992 return NewBB; 993 }; 994 995 // We skip cloning blocks when they have a dominating succ that is not the 996 // succ we are cloning for. 997 auto SkipBlock = [&](BasicBlock *BB) { 998 auto It = DominatingSucc.find(BB); 999 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB; 1000 }; 1001 1002 // First, clone the preheader. 1003 auto *ClonedPH = CloneBlock(LoopPH); 1004 1005 // Then clone all the loop blocks, skipping the ones that aren't necessary. 1006 for (auto *LoopBB : L.blocks()) 1007 if (!SkipBlock(LoopBB)) 1008 CloneBlock(LoopBB); 1009 1010 // Split all the loop exit edges so that when we clone the exit blocks, if 1011 // any of the exit blocks are *also* a preheader for some other loop, we 1012 // don't create multiple predecessors entering the loop header. 1013 for (auto *ExitBB : ExitBlocks) { 1014 if (SkipBlock(ExitBB)) 1015 continue; 1016 1017 // When we are going to clone an exit, we don't need to clone all the 1018 // instructions in the exit block and we want to ensure we have an easy 1019 // place to merge the CFG, so split the exit first. This is always safe to 1020 // do because there cannot be any non-loop predecessors of a loop exit in 1021 // loop simplified form. 1022 auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU); 1023 1024 // Rearrange the names to make it easier to write test cases by having the 1025 // exit block carry the suffix rather than the merge block carrying the 1026 // suffix. 1027 MergeBB->takeName(ExitBB); 1028 ExitBB->setName(Twine(MergeBB->getName()) + ".split"); 1029 1030 // Now clone the original exit block. 1031 auto *ClonedExitBB = CloneBlock(ExitBB); 1032 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && 1033 "Exit block should have been split to have one successor!"); 1034 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && 1035 "Cloned exit block has the wrong successor!"); 1036 1037 // Remap any cloned instructions and create a merge phi node for them. 1038 for (auto ZippedInsts : llvm::zip_first( 1039 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())), 1040 llvm::make_range(ClonedExitBB->begin(), 1041 std::prev(ClonedExitBB->end())))) { 1042 Instruction &I = std::get<0>(ZippedInsts); 1043 Instruction &ClonedI = std::get<1>(ZippedInsts); 1044 1045 // The only instructions in the exit block should be PHI nodes and 1046 // potentially a landing pad. 1047 assert( 1048 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && 1049 "Bad instruction in exit block!"); 1050 // We should have a value map between the instruction and its clone. 1051 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!"); 1052 1053 auto *MergePN = 1054 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi", 1055 &*MergeBB->getFirstInsertionPt()); 1056 I.replaceAllUsesWith(MergePN); 1057 MergePN->addIncoming(&I, ExitBB); 1058 MergePN->addIncoming(&ClonedI, ClonedExitBB); 1059 } 1060 } 1061 1062 // Rewrite the instructions in the cloned blocks to refer to the instructions 1063 // in the cloned blocks. We have to do this as a second pass so that we have 1064 // everything available. Also, we have inserted new instructions which may 1065 // include assume intrinsics, so we update the assumption cache while 1066 // processing this. 1067 for (auto *ClonedBB : NewBlocks) 1068 for (Instruction &I : *ClonedBB) { 1069 RemapInstruction(&I, VMap, 1070 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1071 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 1072 if (II->getIntrinsicID() == Intrinsic::assume) 1073 AC.registerAssumption(II); 1074 } 1075 1076 // Update any PHI nodes in the cloned successors of the skipped blocks to not 1077 // have spurious incoming values. 1078 for (auto *LoopBB : L.blocks()) 1079 if (SkipBlock(LoopBB)) 1080 for (auto *SuccBB : successors(LoopBB)) 1081 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB))) 1082 for (PHINode &PN : ClonedSuccBB->phis()) 1083 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false); 1084 1085 // Remove the cloned parent as a predecessor of any successor we ended up 1086 // cloning other than the unswitched one. 1087 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB)); 1088 for (auto *SuccBB : successors(ParentBB)) { 1089 if (SuccBB == UnswitchedSuccBB) 1090 continue; 1091 1092 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)); 1093 if (!ClonedSuccBB) 1094 continue; 1095 1096 ClonedSuccBB->removePredecessor(ClonedParentBB, 1097 /*KeepOneInputPHIs*/ true); 1098 } 1099 1100 // Replace the cloned branch with an unconditional branch to the cloned 1101 // unswitched successor. 1102 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB)); 1103 ClonedParentBB->getTerminator()->eraseFromParent(); 1104 BranchInst::Create(ClonedSuccBB, ClonedParentBB); 1105 1106 // If there are duplicate entries in the PHI nodes because of multiple edges 1107 // to the unswitched successor, we need to nuke all but one as we replaced it 1108 // with a direct branch. 1109 for (PHINode &PN : ClonedSuccBB->phis()) { 1110 bool Found = false; 1111 // Loop over the incoming operands backwards so we can easily delete as we 1112 // go without invalidating the index. 1113 for (int i = PN.getNumOperands() - 1; i >= 0; --i) { 1114 if (PN.getIncomingBlock(i) != ClonedParentBB) 1115 continue; 1116 if (!Found) { 1117 Found = true; 1118 continue; 1119 } 1120 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false); 1121 } 1122 } 1123 1124 // Record the domtree updates for the new blocks. 1125 SmallPtrSet<BasicBlock *, 4> SuccSet; 1126 for (auto *ClonedBB : NewBlocks) { 1127 for (auto *SuccBB : successors(ClonedBB)) 1128 if (SuccSet.insert(SuccBB).second) 1129 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB}); 1130 SuccSet.clear(); 1131 } 1132 1133 return ClonedPH; 1134 } 1135 1136 /// Recursively clone the specified loop and all of its children. 1137 /// 1138 /// The target parent loop for the clone should be provided, or can be null if 1139 /// the clone is a top-level loop. While cloning, all the blocks are mapped 1140 /// with the provided value map. The entire original loop must be present in 1141 /// the value map. The cloned loop is returned. 1142 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, 1143 const ValueToValueMapTy &VMap, LoopInfo &LI) { 1144 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) { 1145 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!"); 1146 ClonedL.reserveBlocks(OrigL.getNumBlocks()); 1147 for (auto *BB : OrigL.blocks()) { 1148 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB)); 1149 ClonedL.addBlockEntry(ClonedBB); 1150 if (LI.getLoopFor(BB) == &OrigL) 1151 LI.changeLoopFor(ClonedBB, &ClonedL); 1152 } 1153 }; 1154 1155 // We specially handle the first loop because it may get cloned into 1156 // a different parent and because we most commonly are cloning leaf loops. 1157 Loop *ClonedRootL = LI.AllocateLoop(); 1158 if (RootParentL) 1159 RootParentL->addChildLoop(ClonedRootL); 1160 else 1161 LI.addTopLevelLoop(ClonedRootL); 1162 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL); 1163 1164 if (OrigRootL.empty()) 1165 return ClonedRootL; 1166 1167 // If we have a nest, we can quickly clone the entire loop nest using an 1168 // iterative approach because it is a tree. We keep the cloned parent in the 1169 // data structure to avoid repeatedly querying through a map to find it. 1170 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone; 1171 // Build up the loops to clone in reverse order as we'll clone them from the 1172 // back. 1173 for (Loop *ChildL : llvm::reverse(OrigRootL)) 1174 LoopsToClone.push_back({ClonedRootL, ChildL}); 1175 do { 1176 Loop *ClonedParentL, *L; 1177 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val(); 1178 Loop *ClonedL = LI.AllocateLoop(); 1179 ClonedParentL->addChildLoop(ClonedL); 1180 AddClonedBlocksToLoop(*L, *ClonedL); 1181 for (Loop *ChildL : llvm::reverse(*L)) 1182 LoopsToClone.push_back({ClonedL, ChildL}); 1183 } while (!LoopsToClone.empty()); 1184 1185 return ClonedRootL; 1186 } 1187 1188 /// Build the cloned loops of an original loop from unswitching. 1189 /// 1190 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial 1191 /// operation. We need to re-verify that there even is a loop (as the backedge 1192 /// may not have been cloned), and even if there are remaining backedges the 1193 /// backedge set may be different. However, we know that each child loop is 1194 /// undisturbed, we only need to find where to place each child loop within 1195 /// either any parent loop or within a cloned version of the original loop. 1196 /// 1197 /// Because child loops may end up cloned outside of any cloned version of the 1198 /// original loop, multiple cloned sibling loops may be created. All of them 1199 /// are returned so that the newly introduced loop nest roots can be 1200 /// identified. 1201 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks, 1202 const ValueToValueMapTy &VMap, LoopInfo &LI, 1203 SmallVectorImpl<Loop *> &NonChildClonedLoops) { 1204 Loop *ClonedL = nullptr; 1205 1206 auto *OrigPH = OrigL.getLoopPreheader(); 1207 auto *OrigHeader = OrigL.getHeader(); 1208 1209 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH)); 1210 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader)); 1211 1212 // We need to know the loops of the cloned exit blocks to even compute the 1213 // accurate parent loop. If we only clone exits to some parent of the 1214 // original parent, we want to clone into that outer loop. We also keep track 1215 // of the loops that our cloned exit blocks participate in. 1216 Loop *ParentL = nullptr; 1217 SmallVector<BasicBlock *, 4> ClonedExitsInLoops; 1218 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap; 1219 ClonedExitsInLoops.reserve(ExitBlocks.size()); 1220 for (auto *ExitBB : ExitBlocks) 1221 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB))) 1222 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1223 ExitLoopMap[ClonedExitBB] = ExitL; 1224 ClonedExitsInLoops.push_back(ClonedExitBB); 1225 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1226 ParentL = ExitL; 1227 } 1228 assert((!ParentL || ParentL == OrigL.getParentLoop() || 1229 ParentL->contains(OrigL.getParentLoop())) && 1230 "The computed parent loop should always contain (or be) the parent of " 1231 "the original loop."); 1232 1233 // We build the set of blocks dominated by the cloned header from the set of 1234 // cloned blocks out of the original loop. While not all of these will 1235 // necessarily be in the cloned loop, it is enough to establish that they 1236 // aren't in unreachable cycles, etc. 1237 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks; 1238 for (auto *BB : OrigL.blocks()) 1239 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB))) 1240 ClonedLoopBlocks.insert(ClonedBB); 1241 1242 // Rebuild the set of blocks that will end up in the cloned loop. We may have 1243 // skipped cloning some region of this loop which can in turn skip some of 1244 // the backedges so we have to rebuild the blocks in the loop based on the 1245 // backedges that remain after cloning. 1246 SmallVector<BasicBlock *, 16> Worklist; 1247 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop; 1248 for (auto *Pred : predecessors(ClonedHeader)) { 1249 // The only possible non-loop header predecessor is the preheader because 1250 // we know we cloned the loop in simplified form. 1251 if (Pred == ClonedPH) 1252 continue; 1253 1254 // Because the loop was in simplified form, the only non-loop predecessor 1255 // should be the preheader. 1256 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop " 1257 "header other than the preheader " 1258 "that is not part of the loop!"); 1259 1260 // Insert this block into the loop set and on the first visit (and if it 1261 // isn't the header we're currently walking) put it into the worklist to 1262 // recurse through. 1263 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader) 1264 Worklist.push_back(Pred); 1265 } 1266 1267 // If we had any backedges then there *is* a cloned loop. Put the header into 1268 // the loop set and then walk the worklist backwards to find all the blocks 1269 // that remain within the loop after cloning. 1270 if (!BlocksInClonedLoop.empty()) { 1271 BlocksInClonedLoop.insert(ClonedHeader); 1272 1273 while (!Worklist.empty()) { 1274 BasicBlock *BB = Worklist.pop_back_val(); 1275 assert(BlocksInClonedLoop.count(BB) && 1276 "Didn't put block into the loop set!"); 1277 1278 // Insert any predecessors that are in the possible set into the cloned 1279 // set, and if the insert is successful, add them to the worklist. Note 1280 // that we filter on the blocks that are definitely reachable via the 1281 // backedge to the loop header so we may prune out dead code within the 1282 // cloned loop. 1283 for (auto *Pred : predecessors(BB)) 1284 if (ClonedLoopBlocks.count(Pred) && 1285 BlocksInClonedLoop.insert(Pred).second) 1286 Worklist.push_back(Pred); 1287 } 1288 1289 ClonedL = LI.AllocateLoop(); 1290 if (ParentL) { 1291 ParentL->addBasicBlockToLoop(ClonedPH, LI); 1292 ParentL->addChildLoop(ClonedL); 1293 } else { 1294 LI.addTopLevelLoop(ClonedL); 1295 } 1296 NonChildClonedLoops.push_back(ClonedL); 1297 1298 ClonedL->reserveBlocks(BlocksInClonedLoop.size()); 1299 // We don't want to just add the cloned loop blocks based on how we 1300 // discovered them. The original order of blocks was carefully built in 1301 // a way that doesn't rely on predecessor ordering. Rather than re-invent 1302 // that logic, we just re-walk the original blocks (and those of the child 1303 // loops) and filter them as we add them into the cloned loop. 1304 for (auto *BB : OrigL.blocks()) { 1305 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)); 1306 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB)) 1307 continue; 1308 1309 // Directly add the blocks that are only in this loop. 1310 if (LI.getLoopFor(BB) == &OrigL) { 1311 ClonedL->addBasicBlockToLoop(ClonedBB, LI); 1312 continue; 1313 } 1314 1315 // We want to manually add it to this loop and parents. 1316 // Registering it with LoopInfo will happen when we clone the top 1317 // loop for this block. 1318 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop()) 1319 PL->addBlockEntry(ClonedBB); 1320 } 1321 1322 // Now add each child loop whose header remains within the cloned loop. All 1323 // of the blocks within the loop must satisfy the same constraints as the 1324 // header so once we pass the header checks we can just clone the entire 1325 // child loop nest. 1326 for (Loop *ChildL : OrigL) { 1327 auto *ClonedChildHeader = 1328 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1329 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader)) 1330 continue; 1331 1332 #ifndef NDEBUG 1333 // We should never have a cloned child loop header but fail to have 1334 // all of the blocks for that child loop. 1335 for (auto *ChildLoopBB : ChildL->blocks()) 1336 assert(BlocksInClonedLoop.count( 1337 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && 1338 "Child cloned loop has a header within the cloned outer " 1339 "loop but not all of its blocks!"); 1340 #endif 1341 1342 cloneLoopNest(*ChildL, ClonedL, VMap, LI); 1343 } 1344 } 1345 1346 // Now that we've handled all the components of the original loop that were 1347 // cloned into a new loop, we still need to handle anything from the original 1348 // loop that wasn't in a cloned loop. 1349 1350 // Figure out what blocks are left to place within any loop nest containing 1351 // the unswitched loop. If we never formed a loop, the cloned PH is one of 1352 // them. 1353 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet; 1354 if (BlocksInClonedLoop.empty()) 1355 UnloopedBlockSet.insert(ClonedPH); 1356 for (auto *ClonedBB : ClonedLoopBlocks) 1357 if (!BlocksInClonedLoop.count(ClonedBB)) 1358 UnloopedBlockSet.insert(ClonedBB); 1359 1360 // Copy the cloned exits and sort them in ascending loop depth, we'll work 1361 // backwards across these to process them inside out. The order shouldn't 1362 // matter as we're just trying to build up the map from inside-out; we use 1363 // the map in a more stably ordered way below. 1364 auto OrderedClonedExitsInLoops = ClonedExitsInLoops; 1365 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { 1366 return ExitLoopMap.lookup(LHS)->getLoopDepth() < 1367 ExitLoopMap.lookup(RHS)->getLoopDepth(); 1368 }); 1369 1370 // Populate the existing ExitLoopMap with everything reachable from each 1371 // exit, starting from the inner most exit. 1372 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) { 1373 assert(Worklist.empty() && "Didn't clear worklist!"); 1374 1375 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val(); 1376 Loop *ExitL = ExitLoopMap.lookup(ExitBB); 1377 1378 // Walk the CFG back until we hit the cloned PH adding everything reachable 1379 // and in the unlooped set to this exit block's loop. 1380 Worklist.push_back(ExitBB); 1381 do { 1382 BasicBlock *BB = Worklist.pop_back_val(); 1383 // We can stop recursing at the cloned preheader (if we get there). 1384 if (BB == ClonedPH) 1385 continue; 1386 1387 for (BasicBlock *PredBB : predecessors(BB)) { 1388 // If this pred has already been moved to our set or is part of some 1389 // (inner) loop, no update needed. 1390 if (!UnloopedBlockSet.erase(PredBB)) { 1391 assert( 1392 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && 1393 "Predecessor not mapped to a loop!"); 1394 continue; 1395 } 1396 1397 // We just insert into the loop set here. We'll add these blocks to the 1398 // exit loop after we build up the set in an order that doesn't rely on 1399 // predecessor order (which in turn relies on use list order). 1400 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second; 1401 (void)Inserted; 1402 assert(Inserted && "Should only visit an unlooped block once!"); 1403 1404 // And recurse through to its predecessors. 1405 Worklist.push_back(PredBB); 1406 } 1407 } while (!Worklist.empty()); 1408 } 1409 1410 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned 1411 // blocks to their outer loops, walk the cloned blocks and the cloned exits 1412 // in their original order adding them to the correct loop. 1413 1414 // We need a stable insertion order. We use the order of the original loop 1415 // order and map into the correct parent loop. 1416 for (auto *BB : llvm::concat<BasicBlock *const>( 1417 makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops)) 1418 if (Loop *OuterL = ExitLoopMap.lookup(BB)) 1419 OuterL->addBasicBlockToLoop(BB, LI); 1420 1421 #ifndef NDEBUG 1422 for (auto &BBAndL : ExitLoopMap) { 1423 auto *BB = BBAndL.first; 1424 auto *OuterL = BBAndL.second; 1425 assert(LI.getLoopFor(BB) == OuterL && 1426 "Failed to put all blocks into outer loops!"); 1427 } 1428 #endif 1429 1430 // Now that all the blocks are placed into the correct containing loop in the 1431 // absence of child loops, find all the potentially cloned child loops and 1432 // clone them into whatever outer loop we placed their header into. 1433 for (Loop *ChildL : OrigL) { 1434 auto *ClonedChildHeader = 1435 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1436 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader)) 1437 continue; 1438 1439 #ifndef NDEBUG 1440 for (auto *ChildLoopBB : ChildL->blocks()) 1441 assert(VMap.count(ChildLoopBB) && 1442 "Cloned a child loop header but not all of that loops blocks!"); 1443 #endif 1444 1445 NonChildClonedLoops.push_back(cloneLoopNest( 1446 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI)); 1447 } 1448 } 1449 1450 static void 1451 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1452 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, 1453 DominatorTree &DT, MemorySSAUpdater *MSSAU) { 1454 // Find all the dead clones, and remove them from their successors. 1455 SmallVector<BasicBlock *, 16> DeadBlocks; 1456 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks)) 1457 for (auto &VMap : VMaps) 1458 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB))) 1459 if (!DT.isReachableFromEntry(ClonedBB)) { 1460 for (BasicBlock *SuccBB : successors(ClonedBB)) 1461 SuccBB->removePredecessor(ClonedBB); 1462 DeadBlocks.push_back(ClonedBB); 1463 } 1464 1465 // Remove all MemorySSA in the dead blocks 1466 if (MSSAU) { 1467 SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(), 1468 DeadBlocks.end()); 1469 MSSAU->removeBlocks(DeadBlockSet); 1470 } 1471 1472 // Drop any remaining references to break cycles. 1473 for (BasicBlock *BB : DeadBlocks) 1474 BB->dropAllReferences(); 1475 // Erase them from the IR. 1476 for (BasicBlock *BB : DeadBlocks) 1477 BB->eraseFromParent(); 1478 } 1479 1480 static void deleteDeadBlocksFromLoop(Loop &L, 1481 SmallVectorImpl<BasicBlock *> &ExitBlocks, 1482 DominatorTree &DT, LoopInfo &LI, 1483 MemorySSAUpdater *MSSAU) { 1484 // Find all the dead blocks tied to this loop, and remove them from their 1485 // successors. 1486 SmallSetVector<BasicBlock *, 8> DeadBlockSet; 1487 1488 // Start with loop/exit blocks and get a transitive closure of reachable dead 1489 // blocks. 1490 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(), 1491 ExitBlocks.end()); 1492 DeathCandidates.append(L.blocks().begin(), L.blocks().end()); 1493 while (!DeathCandidates.empty()) { 1494 auto *BB = DeathCandidates.pop_back_val(); 1495 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) { 1496 for (BasicBlock *SuccBB : successors(BB)) { 1497 SuccBB->removePredecessor(BB); 1498 DeathCandidates.push_back(SuccBB); 1499 } 1500 DeadBlockSet.insert(BB); 1501 } 1502 } 1503 1504 // Remove all MemorySSA in the dead blocks 1505 if (MSSAU) 1506 MSSAU->removeBlocks(DeadBlockSet); 1507 1508 // Filter out the dead blocks from the exit blocks list so that it can be 1509 // used in the caller. 1510 llvm::erase_if(ExitBlocks, 1511 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1512 1513 // Walk from this loop up through its parents removing all of the dead blocks. 1514 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) { 1515 for (auto *BB : DeadBlockSet) 1516 ParentL->getBlocksSet().erase(BB); 1517 llvm::erase_if(ParentL->getBlocksVector(), 1518 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1519 } 1520 1521 // Now delete the dead child loops. This raw delete will clear them 1522 // recursively. 1523 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) { 1524 if (!DeadBlockSet.count(ChildL->getHeader())) 1525 return false; 1526 1527 assert(llvm::all_of(ChildL->blocks(), 1528 [&](BasicBlock *ChildBB) { 1529 return DeadBlockSet.count(ChildBB); 1530 }) && 1531 "If the child loop header is dead all blocks in the child loop must " 1532 "be dead as well!"); 1533 LI.destroy(ChildL); 1534 return true; 1535 }); 1536 1537 // Remove the loop mappings for the dead blocks and drop all the references 1538 // from these blocks to others to handle cyclic references as we start 1539 // deleting the blocks themselves. 1540 for (auto *BB : DeadBlockSet) { 1541 // Check that the dominator tree has already been updated. 1542 assert(!DT.getNode(BB) && "Should already have cleared domtree!"); 1543 LI.changeLoopFor(BB, nullptr); 1544 BB->dropAllReferences(); 1545 } 1546 1547 // Actually delete the blocks now that they've been fully unhooked from the 1548 // IR. 1549 for (auto *BB : DeadBlockSet) 1550 BB->eraseFromParent(); 1551 } 1552 1553 /// Recompute the set of blocks in a loop after unswitching. 1554 /// 1555 /// This walks from the original headers predecessors to rebuild the loop. We 1556 /// take advantage of the fact that new blocks can't have been added, and so we 1557 /// filter by the original loop's blocks. This also handles potentially 1558 /// unreachable code that we don't want to explore but might be found examining 1559 /// the predecessors of the header. 1560 /// 1561 /// If the original loop is no longer a loop, this will return an empty set. If 1562 /// it remains a loop, all the blocks within it will be added to the set 1563 /// (including those blocks in inner loops). 1564 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L, 1565 LoopInfo &LI) { 1566 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet; 1567 1568 auto *PH = L.getLoopPreheader(); 1569 auto *Header = L.getHeader(); 1570 1571 // A worklist to use while walking backwards from the header. 1572 SmallVector<BasicBlock *, 16> Worklist; 1573 1574 // First walk the predecessors of the header to find the backedges. This will 1575 // form the basis of our walk. 1576 for (auto *Pred : predecessors(Header)) { 1577 // Skip the preheader. 1578 if (Pred == PH) 1579 continue; 1580 1581 // Because the loop was in simplified form, the only non-loop predecessor 1582 // is the preheader. 1583 assert(L.contains(Pred) && "Found a predecessor of the loop header other " 1584 "than the preheader that is not part of the " 1585 "loop!"); 1586 1587 // Insert this block into the loop set and on the first visit and, if it 1588 // isn't the header we're currently walking, put it into the worklist to 1589 // recurse through. 1590 if (LoopBlockSet.insert(Pred).second && Pred != Header) 1591 Worklist.push_back(Pred); 1592 } 1593 1594 // If no backedges were found, we're done. 1595 if (LoopBlockSet.empty()) 1596 return LoopBlockSet; 1597 1598 // We found backedges, recurse through them to identify the loop blocks. 1599 while (!Worklist.empty()) { 1600 BasicBlock *BB = Worklist.pop_back_val(); 1601 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!"); 1602 1603 // No need to walk past the header. 1604 if (BB == Header) 1605 continue; 1606 1607 // Because we know the inner loop structure remains valid we can use the 1608 // loop structure to jump immediately across the entire nested loop. 1609 // Further, because it is in loop simplified form, we can directly jump 1610 // to its preheader afterward. 1611 if (Loop *InnerL = LI.getLoopFor(BB)) 1612 if (InnerL != &L) { 1613 assert(L.contains(InnerL) && 1614 "Should not reach a loop *outside* this loop!"); 1615 // The preheader is the only possible predecessor of the loop so 1616 // insert it into the set and check whether it was already handled. 1617 auto *InnerPH = InnerL->getLoopPreheader(); 1618 assert(L.contains(InnerPH) && "Cannot contain an inner loop block " 1619 "but not contain the inner loop " 1620 "preheader!"); 1621 if (!LoopBlockSet.insert(InnerPH).second) 1622 // The only way to reach the preheader is through the loop body 1623 // itself so if it has been visited the loop is already handled. 1624 continue; 1625 1626 // Insert all of the blocks (other than those already present) into 1627 // the loop set. We expect at least the block that led us to find the 1628 // inner loop to be in the block set, but we may also have other loop 1629 // blocks if they were already enqueued as predecessors of some other 1630 // outer loop block. 1631 for (auto *InnerBB : InnerL->blocks()) { 1632 if (InnerBB == BB) { 1633 assert(LoopBlockSet.count(InnerBB) && 1634 "Block should already be in the set!"); 1635 continue; 1636 } 1637 1638 LoopBlockSet.insert(InnerBB); 1639 } 1640 1641 // Add the preheader to the worklist so we will continue past the 1642 // loop body. 1643 Worklist.push_back(InnerPH); 1644 continue; 1645 } 1646 1647 // Insert any predecessors that were in the original loop into the new 1648 // set, and if the insert is successful, add them to the worklist. 1649 for (auto *Pred : predecessors(BB)) 1650 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second) 1651 Worklist.push_back(Pred); 1652 } 1653 1654 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!"); 1655 1656 // We've found all the blocks participating in the loop, return our completed 1657 // set. 1658 return LoopBlockSet; 1659 } 1660 1661 /// Rebuild a loop after unswitching removes some subset of blocks and edges. 1662 /// 1663 /// The removal may have removed some child loops entirely but cannot have 1664 /// disturbed any remaining child loops. However, they may need to be hoisted 1665 /// to the parent loop (or to be top-level loops). The original loop may be 1666 /// completely removed. 1667 /// 1668 /// The sibling loops resulting from this update are returned. If the original 1669 /// loop remains a valid loop, it will be the first entry in this list with all 1670 /// of the newly sibling loops following it. 1671 /// 1672 /// Returns true if the loop remains a loop after unswitching, and false if it 1673 /// is no longer a loop after unswitching (and should not continue to be 1674 /// referenced). 1675 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1676 LoopInfo &LI, 1677 SmallVectorImpl<Loop *> &HoistedLoops) { 1678 auto *PH = L.getLoopPreheader(); 1679 1680 // Compute the actual parent loop from the exit blocks. Because we may have 1681 // pruned some exits the loop may be different from the original parent. 1682 Loop *ParentL = nullptr; 1683 SmallVector<Loop *, 4> ExitLoops; 1684 SmallVector<BasicBlock *, 4> ExitsInLoops; 1685 ExitsInLoops.reserve(ExitBlocks.size()); 1686 for (auto *ExitBB : ExitBlocks) 1687 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1688 ExitLoops.push_back(ExitL); 1689 ExitsInLoops.push_back(ExitBB); 1690 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1691 ParentL = ExitL; 1692 } 1693 1694 // Recompute the blocks participating in this loop. This may be empty if it 1695 // is no longer a loop. 1696 auto LoopBlockSet = recomputeLoopBlockSet(L, LI); 1697 1698 // If we still have a loop, we need to re-set the loop's parent as the exit 1699 // block set changing may have moved it within the loop nest. Note that this 1700 // can only happen when this loop has a parent as it can only hoist the loop 1701 // *up* the nest. 1702 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) { 1703 // Remove this loop's (original) blocks from all of the intervening loops. 1704 for (Loop *IL = L.getParentLoop(); IL != ParentL; 1705 IL = IL->getParentLoop()) { 1706 IL->getBlocksSet().erase(PH); 1707 for (auto *BB : L.blocks()) 1708 IL->getBlocksSet().erase(BB); 1709 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) { 1710 return BB == PH || L.contains(BB); 1711 }); 1712 } 1713 1714 LI.changeLoopFor(PH, ParentL); 1715 L.getParentLoop()->removeChildLoop(&L); 1716 if (ParentL) 1717 ParentL->addChildLoop(&L); 1718 else 1719 LI.addTopLevelLoop(&L); 1720 } 1721 1722 // Now we update all the blocks which are no longer within the loop. 1723 auto &Blocks = L.getBlocksVector(); 1724 auto BlocksSplitI = 1725 LoopBlockSet.empty() 1726 ? Blocks.begin() 1727 : std::stable_partition( 1728 Blocks.begin(), Blocks.end(), 1729 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); }); 1730 1731 // Before we erase the list of unlooped blocks, build a set of them. 1732 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end()); 1733 if (LoopBlockSet.empty()) 1734 UnloopedBlocks.insert(PH); 1735 1736 // Now erase these blocks from the loop. 1737 for (auto *BB : make_range(BlocksSplitI, Blocks.end())) 1738 L.getBlocksSet().erase(BB); 1739 Blocks.erase(BlocksSplitI, Blocks.end()); 1740 1741 // Sort the exits in ascending loop depth, we'll work backwards across these 1742 // to process them inside out. 1743 llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { 1744 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS); 1745 }); 1746 1747 // We'll build up a set for each exit loop. 1748 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks; 1749 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop. 1750 1751 auto RemoveUnloopedBlocksFromLoop = 1752 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) { 1753 for (auto *BB : UnloopedBlocks) 1754 L.getBlocksSet().erase(BB); 1755 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) { 1756 return UnloopedBlocks.count(BB); 1757 }); 1758 }; 1759 1760 SmallVector<BasicBlock *, 16> Worklist; 1761 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) { 1762 assert(Worklist.empty() && "Didn't clear worklist!"); 1763 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!"); 1764 1765 // Grab the next exit block, in decreasing loop depth order. 1766 BasicBlock *ExitBB = ExitsInLoops.pop_back_val(); 1767 Loop &ExitL = *LI.getLoopFor(ExitBB); 1768 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!"); 1769 1770 // Erase all of the unlooped blocks from the loops between the previous 1771 // exit loop and this exit loop. This works because the ExitInLoops list is 1772 // sorted in increasing order of loop depth and thus we visit loops in 1773 // decreasing order of loop depth. 1774 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop()) 1775 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1776 1777 // Walk the CFG back until we hit the cloned PH adding everything reachable 1778 // and in the unlooped set to this exit block's loop. 1779 Worklist.push_back(ExitBB); 1780 do { 1781 BasicBlock *BB = Worklist.pop_back_val(); 1782 // We can stop recursing at the cloned preheader (if we get there). 1783 if (BB == PH) 1784 continue; 1785 1786 for (BasicBlock *PredBB : predecessors(BB)) { 1787 // If this pred has already been moved to our set or is part of some 1788 // (inner) loop, no update needed. 1789 if (!UnloopedBlocks.erase(PredBB)) { 1790 assert((NewExitLoopBlocks.count(PredBB) || 1791 ExitL.contains(LI.getLoopFor(PredBB))) && 1792 "Predecessor not in a nested loop (or already visited)!"); 1793 continue; 1794 } 1795 1796 // We just insert into the loop set here. We'll add these blocks to the 1797 // exit loop after we build up the set in a deterministic order rather 1798 // than the predecessor-influenced visit order. 1799 bool Inserted = NewExitLoopBlocks.insert(PredBB).second; 1800 (void)Inserted; 1801 assert(Inserted && "Should only visit an unlooped block once!"); 1802 1803 // And recurse through to its predecessors. 1804 Worklist.push_back(PredBB); 1805 } 1806 } while (!Worklist.empty()); 1807 1808 // If blocks in this exit loop were directly part of the original loop (as 1809 // opposed to a child loop) update the map to point to this exit loop. This 1810 // just updates a map and so the fact that the order is unstable is fine. 1811 for (auto *BB : NewExitLoopBlocks) 1812 if (Loop *BBL = LI.getLoopFor(BB)) 1813 if (BBL == &L || !L.contains(BBL)) 1814 LI.changeLoopFor(BB, &ExitL); 1815 1816 // We will remove the remaining unlooped blocks from this loop in the next 1817 // iteration or below. 1818 NewExitLoopBlocks.clear(); 1819 } 1820 1821 // Any remaining unlooped blocks are no longer part of any loop unless they 1822 // are part of some child loop. 1823 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop()) 1824 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1825 for (auto *BB : UnloopedBlocks) 1826 if (Loop *BBL = LI.getLoopFor(BB)) 1827 if (BBL == &L || !L.contains(BBL)) 1828 LI.changeLoopFor(BB, nullptr); 1829 1830 // Sink all the child loops whose headers are no longer in the loop set to 1831 // the parent (or to be top level loops). We reach into the loop and directly 1832 // update its subloop vector to make this batch update efficient. 1833 auto &SubLoops = L.getSubLoopsVector(); 1834 auto SubLoopsSplitI = 1835 LoopBlockSet.empty() 1836 ? SubLoops.begin() 1837 : std::stable_partition( 1838 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) { 1839 return LoopBlockSet.count(SubL->getHeader()); 1840 }); 1841 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) { 1842 HoistedLoops.push_back(HoistedL); 1843 HoistedL->setParentLoop(nullptr); 1844 1845 // To compute the new parent of this hoisted loop we look at where we 1846 // placed the preheader above. We can't lookup the header itself because we 1847 // retained the mapping from the header to the hoisted loop. But the 1848 // preheader and header should have the exact same new parent computed 1849 // based on the set of exit blocks from the original loop as the preheader 1850 // is a predecessor of the header and so reached in the reverse walk. And 1851 // because the loops were all in simplified form the preheader of the 1852 // hoisted loop can't be part of some *other* loop. 1853 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader())) 1854 NewParentL->addChildLoop(HoistedL); 1855 else 1856 LI.addTopLevelLoop(HoistedL); 1857 } 1858 SubLoops.erase(SubLoopsSplitI, SubLoops.end()); 1859 1860 // Actually delete the loop if nothing remained within it. 1861 if (Blocks.empty()) { 1862 assert(SubLoops.empty() && 1863 "Failed to remove all subloops from the original loop!"); 1864 if (Loop *ParentL = L.getParentLoop()) 1865 ParentL->removeChildLoop(llvm::find(*ParentL, &L)); 1866 else 1867 LI.removeLoop(llvm::find(LI, &L)); 1868 LI.destroy(&L); 1869 return false; 1870 } 1871 1872 return true; 1873 } 1874 1875 /// Helper to visit a dominator subtree, invoking a callable on each node. 1876 /// 1877 /// Returning false at any point will stop walking past that node of the tree. 1878 template <typename CallableT> 1879 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) { 1880 SmallVector<DomTreeNode *, 4> DomWorklist; 1881 DomWorklist.push_back(DT[BB]); 1882 #ifndef NDEBUG 1883 SmallPtrSet<DomTreeNode *, 4> Visited; 1884 Visited.insert(DT[BB]); 1885 #endif 1886 do { 1887 DomTreeNode *N = DomWorklist.pop_back_val(); 1888 1889 // Visit this node. 1890 if (!Callable(N->getBlock())) 1891 continue; 1892 1893 // Accumulate the child nodes. 1894 for (DomTreeNode *ChildN : *N) { 1895 assert(Visited.insert(ChildN).second && 1896 "Cannot visit a node twice when walking a tree!"); 1897 DomWorklist.push_back(ChildN); 1898 } 1899 } while (!DomWorklist.empty()); 1900 } 1901 1902 static void unswitchNontrivialInvariants( 1903 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants, 1904 SmallVectorImpl<BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI, 1905 AssumptionCache &AC, function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB, 1906 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) { 1907 auto *ParentBB = TI.getParent(); 1908 BranchInst *BI = dyn_cast<BranchInst>(&TI); 1909 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI); 1910 1911 // We can only unswitch switches, conditional branches with an invariant 1912 // condition, or combining invariant conditions with an instruction. 1913 assert((SI || (BI && BI->isConditional())) && 1914 "Can only unswitch switches and conditional branch!"); 1915 bool FullUnswitch = SI || BI->getCondition() == Invariants[0]; 1916 if (FullUnswitch) 1917 assert(Invariants.size() == 1 && 1918 "Cannot have other invariants with full unswitching!"); 1919 else 1920 assert(isa<Instruction>(BI->getCondition()) && 1921 "Partial unswitching requires an instruction as the condition!"); 1922 1923 if (MSSAU && VerifyMemorySSA) 1924 MSSAU->getMemorySSA()->verifyMemorySSA(); 1925 1926 // Constant and BBs tracking the cloned and continuing successor. When we are 1927 // unswitching the entire condition, this can just be trivially chosen to 1928 // unswitch towards `true`. However, when we are unswitching a set of 1929 // invariants combined with `and` or `or`, the combining operation determines 1930 // the best direction to unswitch: we want to unswitch the direction that will 1931 // collapse the branch. 1932 bool Direction = true; 1933 int ClonedSucc = 0; 1934 if (!FullUnswitch) { 1935 if (cast<Instruction>(BI->getCondition())->getOpcode() != Instruction::Or) { 1936 assert(cast<Instruction>(BI->getCondition())->getOpcode() == 1937 Instruction::And && 1938 "Only `or` and `and` instructions can combine invariants being " 1939 "unswitched."); 1940 Direction = false; 1941 ClonedSucc = 1; 1942 } 1943 } 1944 1945 BasicBlock *RetainedSuccBB = 1946 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest(); 1947 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs; 1948 if (BI) 1949 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc)); 1950 else 1951 for (auto Case : SI->cases()) 1952 if (Case.getCaseSuccessor() != RetainedSuccBB) 1953 UnswitchedSuccBBs.insert(Case.getCaseSuccessor()); 1954 1955 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) && 1956 "Should not unswitch the same successor we are retaining!"); 1957 1958 // The branch should be in this exact loop. Any inner loop's invariant branch 1959 // should be handled by unswitching that inner loop. The caller of this 1960 // routine should filter out any candidates that remain (but were skipped for 1961 // whatever reason). 1962 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!"); 1963 1964 // Compute the parent loop now before we start hacking on things. 1965 Loop *ParentL = L.getParentLoop(); 1966 // Get blocks in RPO order for MSSA update, before changing the CFG. 1967 LoopBlocksRPO LBRPO(&L); 1968 if (MSSAU) 1969 LBRPO.perform(&LI); 1970 1971 // Compute the outer-most loop containing one of our exit blocks. This is the 1972 // furthest up our loopnest which can be mutated, which we will use below to 1973 // update things. 1974 Loop *OuterExitL = &L; 1975 for (auto *ExitBB : ExitBlocks) { 1976 Loop *NewOuterExitL = LI.getLoopFor(ExitBB); 1977 if (!NewOuterExitL) { 1978 // We exited the entire nest with this block, so we're done. 1979 OuterExitL = nullptr; 1980 break; 1981 } 1982 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL)) 1983 OuterExitL = NewOuterExitL; 1984 } 1985 1986 // At this point, we're definitely going to unswitch something so invalidate 1987 // any cached information in ScalarEvolution for the outer most loop 1988 // containing an exit block and all nested loops. 1989 if (SE) { 1990 if (OuterExitL) 1991 SE->forgetLoop(OuterExitL); 1992 else 1993 SE->forgetTopmostLoop(&L); 1994 } 1995 1996 // If the edge from this terminator to a successor dominates that successor, 1997 // store a map from each block in its dominator subtree to it. This lets us 1998 // tell when cloning for a particular successor if a block is dominated by 1999 // some *other* successor with a single data structure. We use this to 2000 // significantly reduce cloning. 2001 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc; 2002 for (auto *SuccBB : llvm::concat<BasicBlock *const>( 2003 makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs)) 2004 if (SuccBB->getUniquePredecessor() || 2005 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 2006 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB); 2007 })) 2008 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) { 2009 DominatingSucc[BB] = SuccBB; 2010 return true; 2011 }); 2012 2013 // Split the preheader, so that we know that there is a safe place to insert 2014 // the conditional branch. We will change the preheader to have a conditional 2015 // branch on LoopCond. The original preheader will become the split point 2016 // between the unswitched versions, and we will have a new preheader for the 2017 // original loop. 2018 BasicBlock *SplitBB = L.getLoopPreheader(); 2019 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU); 2020 2021 // Keep track of the dominator tree updates needed. 2022 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 2023 2024 // Clone the loop for each unswitched successor. 2025 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; 2026 VMaps.reserve(UnswitchedSuccBBs.size()); 2027 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs; 2028 for (auto *SuccBB : UnswitchedSuccBBs) { 2029 VMaps.emplace_back(new ValueToValueMapTy()); 2030 ClonedPHs[SuccBB] = buildClonedLoopBlocks( 2031 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB, 2032 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU); 2033 } 2034 2035 // The stitching of the branched code back together depends on whether we're 2036 // doing full unswitching or not with the exception that we always want to 2037 // nuke the initial terminator placed in the split block. 2038 SplitBB->getTerminator()->eraseFromParent(); 2039 if (FullUnswitch) { 2040 // Splice the terminator from the original loop and rewrite its 2041 // successors. 2042 SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI); 2043 2044 // Keep a clone of the terminator for MSSA updates. 2045 Instruction *NewTI = TI.clone(); 2046 ParentBB->getInstList().push_back(NewTI); 2047 2048 // First wire up the moved terminator to the preheaders. 2049 if (BI) { 2050 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2051 BI->setSuccessor(ClonedSucc, ClonedPH); 2052 BI->setSuccessor(1 - ClonedSucc, LoopPH); 2053 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2054 } else { 2055 assert(SI && "Must either be a branch or switch!"); 2056 2057 // Walk the cases and directly update their successors. 2058 assert(SI->getDefaultDest() == RetainedSuccBB && 2059 "Not retaining default successor!"); 2060 SI->setDefaultDest(LoopPH); 2061 for (auto &Case : SI->cases()) 2062 if (Case.getCaseSuccessor() == RetainedSuccBB) 2063 Case.setSuccessor(LoopPH); 2064 else 2065 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second); 2066 2067 // We need to use the set to populate domtree updates as even when there 2068 // are multiple cases pointing at the same successor we only want to 2069 // remove and insert one edge in the domtree. 2070 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2071 DTUpdates.push_back( 2072 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second}); 2073 } 2074 2075 if (MSSAU) { 2076 DT.applyUpdates(DTUpdates); 2077 DTUpdates.clear(); 2078 2079 // Remove all but one edge to the retained block and all unswitched 2080 // blocks. This is to avoid having duplicate entries in the cloned Phis, 2081 // when we know we only keep a single edge for each case. 2082 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB); 2083 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2084 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB); 2085 2086 for (auto &VMap : VMaps) 2087 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, 2088 /*IgnoreIncomingWithNoClones=*/true); 2089 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); 2090 2091 // Remove all edges to unswitched blocks. 2092 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2093 MSSAU->removeEdge(ParentBB, SuccBB); 2094 } 2095 2096 // Now unhook the successor relationship as we'll be replacing 2097 // the terminator with a direct branch. This is much simpler for branches 2098 // than switches so we handle those first. 2099 if (BI) { 2100 // Remove the parent as a predecessor of the unswitched successor. 2101 assert(UnswitchedSuccBBs.size() == 1 && 2102 "Only one possible unswitched block for a branch!"); 2103 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin(); 2104 UnswitchedSuccBB->removePredecessor(ParentBB, 2105 /*KeepOneInputPHIs*/ true); 2106 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB}); 2107 } else { 2108 // Note that we actually want to remove the parent block as a predecessor 2109 // of *every* case successor. The case successor is either unswitched, 2110 // completely eliminating an edge from the parent to that successor, or it 2111 // is a duplicate edge to the retained successor as the retained successor 2112 // is always the default successor and as we'll replace this with a direct 2113 // branch we no longer need the duplicate entries in the PHI nodes. 2114 SwitchInst *NewSI = cast<SwitchInst>(NewTI); 2115 assert(NewSI->getDefaultDest() == RetainedSuccBB && 2116 "Not retaining default successor!"); 2117 for (auto &Case : NewSI->cases()) 2118 Case.getCaseSuccessor()->removePredecessor( 2119 ParentBB, 2120 /*KeepOneInputPHIs*/ true); 2121 2122 // We need to use the set to populate domtree updates as even when there 2123 // are multiple cases pointing at the same successor we only want to 2124 // remove and insert one edge in the domtree. 2125 for (BasicBlock *SuccBB : UnswitchedSuccBBs) 2126 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB}); 2127 } 2128 2129 // After MSSAU update, remove the cloned terminator instruction NewTI. 2130 ParentBB->getTerminator()->eraseFromParent(); 2131 2132 // Create a new unconditional branch to the continuing block (as opposed to 2133 // the one cloned). 2134 BranchInst::Create(RetainedSuccBB, ParentBB); 2135 } else { 2136 assert(BI && "Only branches have partial unswitching."); 2137 assert(UnswitchedSuccBBs.size() == 1 && 2138 "Only one possible unswitched block for a branch!"); 2139 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2140 // When doing a partial unswitch, we have to do a bit more work to build up 2141 // the branch in the split block. 2142 buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction, 2143 *ClonedPH, *LoopPH); 2144 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 2145 2146 if (MSSAU) { 2147 DT.applyUpdates(DTUpdates); 2148 DTUpdates.clear(); 2149 2150 // Perform MSSA cloning updates. 2151 for (auto &VMap : VMaps) 2152 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, 2153 /*IgnoreIncomingWithNoClones=*/true); 2154 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); 2155 } 2156 } 2157 2158 // Apply the updates accumulated above to get an up-to-date dominator tree. 2159 DT.applyUpdates(DTUpdates); 2160 2161 // Now that we have an accurate dominator tree, first delete the dead cloned 2162 // blocks so that we can accurately build any cloned loops. It is important to 2163 // not delete the blocks from the original loop yet because we still want to 2164 // reference the original loop to understand the cloned loop's structure. 2165 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU); 2166 2167 // Build the cloned loop structure itself. This may be substantially 2168 // different from the original structure due to the simplified CFG. This also 2169 // handles inserting all the cloned blocks into the correct loops. 2170 SmallVector<Loop *, 4> NonChildClonedLoops; 2171 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps) 2172 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops); 2173 2174 // Now that our cloned loops have been built, we can update the original loop. 2175 // First we delete the dead blocks from it and then we rebuild the loop 2176 // structure taking these deletions into account. 2177 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU); 2178 2179 if (MSSAU && VerifyMemorySSA) 2180 MSSAU->getMemorySSA()->verifyMemorySSA(); 2181 2182 SmallVector<Loop *, 4> HoistedLoops; 2183 bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops); 2184 2185 if (MSSAU && VerifyMemorySSA) 2186 MSSAU->getMemorySSA()->verifyMemorySSA(); 2187 2188 // This transformation has a high risk of corrupting the dominator tree, and 2189 // the below steps to rebuild loop structures will result in hard to debug 2190 // errors in that case so verify that the dominator tree is sane first. 2191 // FIXME: Remove this when the bugs stop showing up and rely on existing 2192 // verification steps. 2193 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2194 2195 if (BI) { 2196 // If we unswitched a branch which collapses the condition to a known 2197 // constant we want to replace all the uses of the invariants within both 2198 // the original and cloned blocks. We do this here so that we can use the 2199 // now updated dominator tree to identify which side the users are on. 2200 assert(UnswitchedSuccBBs.size() == 1 && 2201 "Only one possible unswitched block for a branch!"); 2202 BasicBlock *ClonedPH = ClonedPHs.begin()->second; 2203 2204 // When considering multiple partially-unswitched invariants 2205 // we cant just go replace them with constants in both branches. 2206 // 2207 // For 'AND' we infer that true branch ("continue") means true 2208 // for each invariant operand. 2209 // For 'OR' we can infer that false branch ("continue") means false 2210 // for each invariant operand. 2211 // So it happens that for multiple-partial case we dont replace 2212 // in the unswitched branch. 2213 bool ReplaceUnswitched = FullUnswitch || (Invariants.size() == 1); 2214 2215 ConstantInt *UnswitchedReplacement = 2216 Direction ? ConstantInt::getTrue(BI->getContext()) 2217 : ConstantInt::getFalse(BI->getContext()); 2218 ConstantInt *ContinueReplacement = 2219 Direction ? ConstantInt::getFalse(BI->getContext()) 2220 : ConstantInt::getTrue(BI->getContext()); 2221 for (Value *Invariant : Invariants) 2222 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); 2223 UI != UE;) { 2224 // Grab the use and walk past it so we can clobber it in the use list. 2225 Use *U = &*UI++; 2226 Instruction *UserI = dyn_cast<Instruction>(U->getUser()); 2227 if (!UserI) 2228 continue; 2229 2230 // Replace it with the 'continue' side if in the main loop body, and the 2231 // unswitched if in the cloned blocks. 2232 if (DT.dominates(LoopPH, UserI->getParent())) 2233 U->set(ContinueReplacement); 2234 else if (ReplaceUnswitched && 2235 DT.dominates(ClonedPH, UserI->getParent())) 2236 U->set(UnswitchedReplacement); 2237 } 2238 } 2239 2240 // We can change which blocks are exit blocks of all the cloned sibling 2241 // loops, the current loop, and any parent loops which shared exit blocks 2242 // with the current loop. As a consequence, we need to re-form LCSSA for 2243 // them. But we shouldn't need to re-form LCSSA for any child loops. 2244 // FIXME: This could be made more efficient by tracking which exit blocks are 2245 // new, and focusing on them, but that isn't likely to be necessary. 2246 // 2247 // In order to reasonably rebuild LCSSA we need to walk inside-out across the 2248 // loop nest and update every loop that could have had its exits changed. We 2249 // also need to cover any intervening loops. We add all of these loops to 2250 // a list and sort them by loop depth to achieve this without updating 2251 // unnecessary loops. 2252 auto UpdateLoop = [&](Loop &UpdateL) { 2253 #ifndef NDEBUG 2254 UpdateL.verifyLoop(); 2255 for (Loop *ChildL : UpdateL) { 2256 ChildL->verifyLoop(); 2257 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) && 2258 "Perturbed a child loop's LCSSA form!"); 2259 } 2260 #endif 2261 // First build LCSSA for this loop so that we can preserve it when 2262 // forming dedicated exits. We don't want to perturb some other loop's 2263 // LCSSA while doing that CFG edit. 2264 formLCSSA(UpdateL, DT, &LI, nullptr); 2265 2266 // For loops reached by this loop's original exit blocks we may 2267 // introduced new, non-dedicated exits. At least try to re-form dedicated 2268 // exits for these loops. This may fail if they couldn't have dedicated 2269 // exits to start with. 2270 formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true); 2271 }; 2272 2273 // For non-child cloned loops and hoisted loops, we just need to update LCSSA 2274 // and we can do it in any order as they don't nest relative to each other. 2275 // 2276 // Also check if any of the loops we have updated have become top-level loops 2277 // as that will necessitate widening the outer loop scope. 2278 for (Loop *UpdatedL : 2279 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) { 2280 UpdateLoop(*UpdatedL); 2281 if (!UpdatedL->getParentLoop()) 2282 OuterExitL = nullptr; 2283 } 2284 if (IsStillLoop) { 2285 UpdateLoop(L); 2286 if (!L.getParentLoop()) 2287 OuterExitL = nullptr; 2288 } 2289 2290 // If the original loop had exit blocks, walk up through the outer most loop 2291 // of those exit blocks to update LCSSA and form updated dedicated exits. 2292 if (OuterExitL != &L) 2293 for (Loop *OuterL = ParentL; OuterL != OuterExitL; 2294 OuterL = OuterL->getParentLoop()) 2295 UpdateLoop(*OuterL); 2296 2297 #ifndef NDEBUG 2298 // Verify the entire loop structure to catch any incorrect updates before we 2299 // progress in the pass pipeline. 2300 LI.verify(DT); 2301 #endif 2302 2303 // Now that we've unswitched something, make callbacks to report the changes. 2304 // For that we need to merge together the updated loops and the cloned loops 2305 // and check whether the original loop survived. 2306 SmallVector<Loop *, 4> SibLoops; 2307 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) 2308 if (UpdatedL->getParentLoop() == ParentL) 2309 SibLoops.push_back(UpdatedL); 2310 UnswitchCB(IsStillLoop, SibLoops); 2311 2312 if (MSSAU && VerifyMemorySSA) 2313 MSSAU->getMemorySSA()->verifyMemorySSA(); 2314 2315 if (BI) 2316 ++NumBranches; 2317 else 2318 ++NumSwitches; 2319 } 2320 2321 /// Recursively compute the cost of a dominator subtree based on the per-block 2322 /// cost map provided. 2323 /// 2324 /// The recursive computation is memozied into the provided DT-indexed cost map 2325 /// to allow querying it for most nodes in the domtree without it becoming 2326 /// quadratic. 2327 static int 2328 computeDomSubtreeCost(DomTreeNode &N, 2329 const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap, 2330 SmallDenseMap<DomTreeNode *, int, 4> &DTCostMap) { 2331 // Don't accumulate cost (or recurse through) blocks not in our block cost 2332 // map and thus not part of the duplication cost being considered. 2333 auto BBCostIt = BBCostMap.find(N.getBlock()); 2334 if (BBCostIt == BBCostMap.end()) 2335 return 0; 2336 2337 // Lookup this node to see if we already computed its cost. 2338 auto DTCostIt = DTCostMap.find(&N); 2339 if (DTCostIt != DTCostMap.end()) 2340 return DTCostIt->second; 2341 2342 // If not, we have to compute it. We can't use insert above and update 2343 // because computing the cost may insert more things into the map. 2344 int Cost = std::accumulate( 2345 N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) { 2346 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap); 2347 }); 2348 bool Inserted = DTCostMap.insert({&N, Cost}).second; 2349 (void)Inserted; 2350 assert(Inserted && "Should not insert a node while visiting children!"); 2351 return Cost; 2352 } 2353 2354 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch, 2355 /// making the following replacement: 2356 /// 2357 /// --code before guard-- 2358 /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ] 2359 /// --code after guard-- 2360 /// 2361 /// into 2362 /// 2363 /// --code before guard-- 2364 /// br i1 %cond, label %guarded, label %deopt 2365 /// 2366 /// guarded: 2367 /// --code after guard-- 2368 /// 2369 /// deopt: 2370 /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ] 2371 /// unreachable 2372 /// 2373 /// It also makes all relevant DT and LI updates, so that all structures are in 2374 /// valid state after this transform. 2375 static BranchInst * 2376 turnGuardIntoBranch(IntrinsicInst *GI, Loop &L, 2377 SmallVectorImpl<BasicBlock *> &ExitBlocks, 2378 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) { 2379 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 2380 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n"); 2381 BasicBlock *CheckBB = GI->getParent(); 2382 2383 if (MSSAU && VerifyMemorySSA) 2384 MSSAU->getMemorySSA()->verifyMemorySSA(); 2385 2386 // Remove all CheckBB's successors from DomTree. A block can be seen among 2387 // successors more than once, but for DomTree it should be added only once. 2388 SmallPtrSet<BasicBlock *, 4> Successors; 2389 for (auto *Succ : successors(CheckBB)) 2390 if (Successors.insert(Succ).second) 2391 DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ}); 2392 2393 Instruction *DeoptBlockTerm = 2394 SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true); 2395 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator()); 2396 // SplitBlockAndInsertIfThen inserts control flow that branches to 2397 // DeoptBlockTerm if the condition is true. We want the opposite. 2398 CheckBI->swapSuccessors(); 2399 2400 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0); 2401 GuardedBlock->setName("guarded"); 2402 CheckBI->getSuccessor(1)->setName("deopt"); 2403 BasicBlock *DeoptBlock = CheckBI->getSuccessor(1); 2404 2405 // We now have a new exit block. 2406 ExitBlocks.push_back(CheckBI->getSuccessor(1)); 2407 2408 if (MSSAU) 2409 MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI); 2410 2411 GI->moveBefore(DeoptBlockTerm); 2412 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext())); 2413 2414 // Add new successors of CheckBB into DomTree. 2415 for (auto *Succ : successors(CheckBB)) 2416 DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ}); 2417 2418 // Now the blocks that used to be CheckBB's successors are GuardedBlock's 2419 // successors. 2420 for (auto *Succ : Successors) 2421 DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ}); 2422 2423 // Make proper changes to DT. 2424 DT.applyUpdates(DTUpdates); 2425 // Inform LI of a new loop block. 2426 L.addBasicBlockToLoop(GuardedBlock, LI); 2427 2428 if (MSSAU) { 2429 MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI)); 2430 MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::End); 2431 if (VerifyMemorySSA) 2432 MSSAU->getMemorySSA()->verifyMemorySSA(); 2433 } 2434 2435 ++NumGuards; 2436 return CheckBI; 2437 } 2438 2439 /// Cost multiplier is a way to limit potentially exponential behavior 2440 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch 2441 /// candidates available. Also accounting for the number of "sibling" loops with 2442 /// the idea to account for previous unswitches that already happened on this 2443 /// cluster of loops. There was an attempt to keep this formula simple, 2444 /// just enough to limit the worst case behavior. Even if it is not that simple 2445 /// now it is still not an attempt to provide a detailed heuristic size 2446 /// prediction. 2447 /// 2448 /// TODO: Make a proper accounting of "explosion" effect for all kinds of 2449 /// unswitch candidates, making adequate predictions instead of wild guesses. 2450 /// That requires knowing not just the number of "remaining" candidates but 2451 /// also costs of unswitching for each of these candidates. 2452 static int calculateUnswitchCostMultiplier( 2453 Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT, 2454 ArrayRef<std::pair<Instruction *, TinyPtrVector<Value *>>> 2455 UnswitchCandidates) { 2456 2457 // Guards and other exiting conditions do not contribute to exponential 2458 // explosion as soon as they dominate the latch (otherwise there might be 2459 // another path to the latch remaining that does not allow to eliminate the 2460 // loop copy on unswitch). 2461 BasicBlock *Latch = L.getLoopLatch(); 2462 BasicBlock *CondBlock = TI.getParent(); 2463 if (DT.dominates(CondBlock, Latch) && 2464 (isGuard(&TI) || 2465 llvm::count_if(successors(&TI), [&L](BasicBlock *SuccBB) { 2466 return L.contains(SuccBB); 2467 }) <= 1)) { 2468 NumCostMultiplierSkipped++; 2469 return 1; 2470 } 2471 2472 auto *ParentL = L.getParentLoop(); 2473 int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size() 2474 : std::distance(LI.begin(), LI.end())); 2475 // Count amount of clones that all the candidates might cause during 2476 // unswitching. Branch/guard counts as 1, switch counts as log2 of its cases. 2477 int UnswitchedClones = 0; 2478 for (auto Candidate : UnswitchCandidates) { 2479 Instruction *CI = Candidate.first; 2480 BasicBlock *CondBlock = CI->getParent(); 2481 bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch); 2482 if (isGuard(CI)) { 2483 if (!SkipExitingSuccessors) 2484 UnswitchedClones++; 2485 continue; 2486 } 2487 int NonExitingSuccessors = llvm::count_if( 2488 successors(CondBlock), [SkipExitingSuccessors, &L](BasicBlock *SuccBB) { 2489 return !SkipExitingSuccessors || L.contains(SuccBB); 2490 }); 2491 UnswitchedClones += Log2_32(NonExitingSuccessors); 2492 } 2493 2494 // Ignore up to the "unscaled candidates" number of unswitch candidates 2495 // when calculating the power-of-two scaling of the cost. The main idea 2496 // with this control is to allow a small number of unswitches to happen 2497 // and rely more on siblings multiplier (see below) when the number 2498 // of candidates is small. 2499 unsigned ClonesPower = 2500 std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0); 2501 2502 // Allowing top-level loops to spread a bit more than nested ones. 2503 int SiblingsMultiplier = 2504 std::max((ParentL ? SiblingsCount 2505 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv), 2506 1); 2507 // Compute the cost multiplier in a way that won't overflow by saturating 2508 // at an upper bound. 2509 int CostMultiplier; 2510 if (ClonesPower > Log2_32(UnswitchThreshold) || 2511 SiblingsMultiplier > UnswitchThreshold) 2512 CostMultiplier = UnswitchThreshold; 2513 else 2514 CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower), 2515 (int)UnswitchThreshold); 2516 2517 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier 2518 << " (siblings " << SiblingsMultiplier << " * clones " 2519 << (1 << ClonesPower) << ")" 2520 << " for unswitch candidate: " << TI << "\n"); 2521 return CostMultiplier; 2522 } 2523 2524 static bool 2525 unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI, 2526 AssumptionCache &AC, TargetTransformInfo &TTI, 2527 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB, 2528 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) { 2529 // Collect all invariant conditions within this loop (as opposed to an inner 2530 // loop which would be handled when visiting that inner loop). 2531 SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4> 2532 UnswitchCandidates; 2533 2534 // Whether or not we should also collect guards in the loop. 2535 bool CollectGuards = false; 2536 if (UnswitchGuards) { 2537 auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction( 2538 Intrinsic::getName(Intrinsic::experimental_guard)); 2539 if (GuardDecl && !GuardDecl->use_empty()) 2540 CollectGuards = true; 2541 } 2542 2543 for (auto *BB : L.blocks()) { 2544 if (LI.getLoopFor(BB) != &L) 2545 continue; 2546 2547 if (CollectGuards) 2548 for (auto &I : *BB) 2549 if (isGuard(&I)) { 2550 auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0); 2551 // TODO: Support AND, OR conditions and partial unswitching. 2552 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond)) 2553 UnswitchCandidates.push_back({&I, {Cond}}); 2554 } 2555 2556 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 2557 // We can only consider fully loop-invariant switch conditions as we need 2558 // to completely eliminate the switch after unswitching. 2559 if (!isa<Constant>(SI->getCondition()) && 2560 L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor()) 2561 UnswitchCandidates.push_back({SI, {SI->getCondition()}}); 2562 continue; 2563 } 2564 2565 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 2566 if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) || 2567 BI->getSuccessor(0) == BI->getSuccessor(1)) 2568 continue; 2569 2570 if (L.isLoopInvariant(BI->getCondition())) { 2571 UnswitchCandidates.push_back({BI, {BI->getCondition()}}); 2572 continue; 2573 } 2574 2575 Instruction &CondI = *cast<Instruction>(BI->getCondition()); 2576 if (CondI.getOpcode() != Instruction::And && 2577 CondI.getOpcode() != Instruction::Or) 2578 continue; 2579 2580 TinyPtrVector<Value *> Invariants = 2581 collectHomogenousInstGraphLoopInvariants(L, CondI, LI); 2582 if (Invariants.empty()) 2583 continue; 2584 2585 UnswitchCandidates.push_back({BI, std::move(Invariants)}); 2586 } 2587 2588 // If we didn't find any candidates, we're done. 2589 if (UnswitchCandidates.empty()) 2590 return false; 2591 2592 // Check if there are irreducible CFG cycles in this loop. If so, we cannot 2593 // easily unswitch non-trivial edges out of the loop. Doing so might turn the 2594 // irreducible control flow into reducible control flow and introduce new 2595 // loops "out of thin air". If we ever discover important use cases for doing 2596 // this, we can add support to loop unswitch, but it is a lot of complexity 2597 // for what seems little or no real world benefit. 2598 LoopBlocksRPO RPOT(&L); 2599 RPOT.perform(&LI); 2600 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 2601 return false; 2602 2603 SmallVector<BasicBlock *, 4> ExitBlocks; 2604 L.getUniqueExitBlocks(ExitBlocks); 2605 2606 // We cannot unswitch if exit blocks contain a cleanuppad instruction as we 2607 // don't know how to split those exit blocks. 2608 // FIXME: We should teach SplitBlock to handle this and remove this 2609 // restriction. 2610 for (auto *ExitBB : ExitBlocks) 2611 if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) { 2612 dbgs() << "Cannot unswitch because of cleanuppad in exit block\n"; 2613 return false; 2614 } 2615 2616 LLVM_DEBUG( 2617 dbgs() << "Considering " << UnswitchCandidates.size() 2618 << " non-trivial loop invariant conditions for unswitching.\n"); 2619 2620 // Given that unswitching these terminators will require duplicating parts of 2621 // the loop, so we need to be able to model that cost. Compute the ephemeral 2622 // values and set up a data structure to hold per-BB costs. We cache each 2623 // block's cost so that we don't recompute this when considering different 2624 // subsets of the loop for duplication during unswitching. 2625 SmallPtrSet<const Value *, 4> EphValues; 2626 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues); 2627 SmallDenseMap<BasicBlock *, int, 4> BBCostMap; 2628 2629 // Compute the cost of each block, as well as the total loop cost. Also, bail 2630 // out if we see instructions which are incompatible with loop unswitching 2631 // (convergent, noduplicate, or cross-basic-block tokens). 2632 // FIXME: We might be able to safely handle some of these in non-duplicated 2633 // regions. 2634 int LoopCost = 0; 2635 for (auto *BB : L.blocks()) { 2636 int Cost = 0; 2637 for (auto &I : *BB) { 2638 if (EphValues.count(&I)) 2639 continue; 2640 2641 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB)) 2642 return false; 2643 if (auto CS = CallSite(&I)) 2644 if (CS.isConvergent() || CS.cannotDuplicate()) 2645 return false; 2646 2647 Cost += TTI.getUserCost(&I); 2648 } 2649 assert(Cost >= 0 && "Must not have negative costs!"); 2650 LoopCost += Cost; 2651 assert(LoopCost >= 0 && "Must not have negative loop costs!"); 2652 BBCostMap[BB] = Cost; 2653 } 2654 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n"); 2655 2656 // Now we find the best candidate by searching for the one with the following 2657 // properties in order: 2658 // 2659 // 1) An unswitching cost below the threshold 2660 // 2) The smallest number of duplicated unswitch candidates (to avoid 2661 // creating redundant subsequent unswitching) 2662 // 3) The smallest cost after unswitching. 2663 // 2664 // We prioritize reducing fanout of unswitch candidates provided the cost 2665 // remains below the threshold because this has a multiplicative effect. 2666 // 2667 // This requires memoizing each dominator subtree to avoid redundant work. 2668 // 2669 // FIXME: Need to actually do the number of candidates part above. 2670 SmallDenseMap<DomTreeNode *, int, 4> DTCostMap; 2671 // Given a terminator which might be unswitched, computes the non-duplicated 2672 // cost for that terminator. 2673 auto ComputeUnswitchedCost = [&](Instruction &TI, bool FullUnswitch) { 2674 BasicBlock &BB = *TI.getParent(); 2675 SmallPtrSet<BasicBlock *, 4> Visited; 2676 2677 int Cost = LoopCost; 2678 for (BasicBlock *SuccBB : successors(&BB)) { 2679 // Don't count successors more than once. 2680 if (!Visited.insert(SuccBB).second) 2681 continue; 2682 2683 // If this is a partial unswitch candidate, then it must be a conditional 2684 // branch with a condition of either `or` or `and`. In that case, one of 2685 // the successors is necessarily duplicated, so don't even try to remove 2686 // its cost. 2687 if (!FullUnswitch) { 2688 auto &BI = cast<BranchInst>(TI); 2689 if (cast<Instruction>(BI.getCondition())->getOpcode() == 2690 Instruction::And) { 2691 if (SuccBB == BI.getSuccessor(1)) 2692 continue; 2693 } else { 2694 assert(cast<Instruction>(BI.getCondition())->getOpcode() == 2695 Instruction::Or && 2696 "Only `and` and `or` conditions can result in a partial " 2697 "unswitch!"); 2698 if (SuccBB == BI.getSuccessor(0)) 2699 continue; 2700 } 2701 } 2702 2703 // This successor's domtree will not need to be duplicated after 2704 // unswitching if the edge to the successor dominates it (and thus the 2705 // entire tree). This essentially means there is no other path into this 2706 // subtree and so it will end up live in only one clone of the loop. 2707 if (SuccBB->getUniquePredecessor() || 2708 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 2709 return PredBB == &BB || DT.dominates(SuccBB, PredBB); 2710 })) { 2711 Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap); 2712 assert(Cost >= 0 && 2713 "Non-duplicated cost should never exceed total loop cost!"); 2714 } 2715 } 2716 2717 // Now scale the cost by the number of unique successors minus one. We 2718 // subtract one because there is already at least one copy of the entire 2719 // loop. This is computing the new cost of unswitching a condition. 2720 // Note that guards always have 2 unique successors that are implicit and 2721 // will be materialized if we decide to unswitch it. 2722 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size(); 2723 assert(SuccessorsCount > 1 && 2724 "Cannot unswitch a condition without multiple distinct successors!"); 2725 return Cost * (SuccessorsCount - 1); 2726 }; 2727 Instruction *BestUnswitchTI = nullptr; 2728 int BestUnswitchCost = 0; 2729 ArrayRef<Value *> BestUnswitchInvariants; 2730 for (auto &TerminatorAndInvariants : UnswitchCandidates) { 2731 Instruction &TI = *TerminatorAndInvariants.first; 2732 ArrayRef<Value *> Invariants = TerminatorAndInvariants.second; 2733 BranchInst *BI = dyn_cast<BranchInst>(&TI); 2734 int CandidateCost = ComputeUnswitchedCost( 2735 TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 && 2736 Invariants[0] == BI->getCondition())); 2737 // Calculate cost multiplier which is a tool to limit potentially 2738 // exponential behavior of loop-unswitch. 2739 if (EnableUnswitchCostMultiplier) { 2740 int CostMultiplier = 2741 calculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates); 2742 assert( 2743 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) && 2744 "cost multiplier needs to be in the range of 1..UnswitchThreshold"); 2745 CandidateCost *= CostMultiplier; 2746 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 2747 << " (multiplier: " << CostMultiplier << ")" 2748 << " for unswitch candidate: " << TI << "\n"); 2749 } else { 2750 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost 2751 << " for unswitch candidate: " << TI << "\n"); 2752 } 2753 2754 if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) { 2755 BestUnswitchTI = &TI; 2756 BestUnswitchCost = CandidateCost; 2757 BestUnswitchInvariants = Invariants; 2758 } 2759 } 2760 assert(BestUnswitchTI && "Failed to find loop unswitch candidate"); 2761 2762 if (BestUnswitchCost >= UnswitchThreshold) { 2763 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " 2764 << BestUnswitchCost << "\n"); 2765 return false; 2766 } 2767 2768 // If the best candidate is a guard, turn it into a branch. 2769 if (isGuard(BestUnswitchTI)) 2770 BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L, 2771 ExitBlocks, DT, LI, MSSAU); 2772 2773 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = " 2774 << BestUnswitchCost << ") terminator: " << *BestUnswitchTI 2775 << "\n"); 2776 unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants, 2777 ExitBlocks, DT, LI, AC, UnswitchCB, SE, MSSAU); 2778 return true; 2779 } 2780 2781 /// Unswitch control flow predicated on loop invariant conditions. 2782 /// 2783 /// This first hoists all branches or switches which are trivial (IE, do not 2784 /// require duplicating any part of the loop) out of the loop body. It then 2785 /// looks at other loop invariant control flows and tries to unswitch those as 2786 /// well by cloning the loop if the result is small enough. 2787 /// 2788 /// The `DT`, `LI`, `AC`, `TTI` parameters are required analyses that are also 2789 /// updated based on the unswitch. 2790 /// The `MSSA` analysis is also updated if valid (i.e. its use is enabled). 2791 /// 2792 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is 2793 /// true, we will attempt to do non-trivial unswitching as well as trivial 2794 /// unswitching. 2795 /// 2796 /// The `UnswitchCB` callback provided will be run after unswitching is 2797 /// complete, with the first parameter set to `true` if the provided loop 2798 /// remains a loop, and a list of new sibling loops created. 2799 /// 2800 /// If `SE` is non-null, we will update that analysis based on the unswitching 2801 /// done. 2802 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, 2803 AssumptionCache &AC, TargetTransformInfo &TTI, 2804 bool NonTrivial, 2805 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB, 2806 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) { 2807 assert(L.isRecursivelyLCSSAForm(DT, LI) && 2808 "Loops must be in LCSSA form before unswitching."); 2809 bool Changed = false; 2810 2811 // Must be in loop simplified form: we need a preheader and dedicated exits. 2812 if (!L.isLoopSimplifyForm()) 2813 return false; 2814 2815 // Try trivial unswitch first before loop over other basic blocks in the loop. 2816 if (unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) { 2817 // If we unswitched successfully we will want to clean up the loop before 2818 // processing it further so just mark it as unswitched and return. 2819 UnswitchCB(/*CurrentLoopValid*/ true, {}); 2820 return true; 2821 } 2822 2823 // If we're not doing non-trivial unswitching, we're done. We both accept 2824 // a parameter but also check a local flag that can be used for testing 2825 // a debugging. 2826 if (!NonTrivial && !EnableNonTrivialUnswitch) 2827 return false; 2828 2829 // For non-trivial unswitching, because it often creates new loops, we rely on 2830 // the pass manager to iterate on the loops rather than trying to immediately 2831 // reach a fixed point. There is no substantial advantage to iterating 2832 // internally, and if any of the new loops are simplified enough to contain 2833 // trivial unswitching we want to prefer those. 2834 2835 // Try to unswitch the best invariant condition. We prefer this full unswitch to 2836 // a partial unswitch when possible below the threshold. 2837 if (unswitchBestCondition(L, DT, LI, AC, TTI, UnswitchCB, SE, MSSAU)) 2838 return true; 2839 2840 // No other opportunities to unswitch. 2841 return Changed; 2842 } 2843 2844 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, 2845 LoopStandardAnalysisResults &AR, 2846 LPMUpdater &U) { 2847 Function &F = *L.getHeader()->getParent(); 2848 (void)F; 2849 2850 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L 2851 << "\n"); 2852 2853 // Save the current loop name in a variable so that we can report it even 2854 // after it has been deleted. 2855 std::string LoopName = L.getName(); 2856 2857 auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid, 2858 ArrayRef<Loop *> NewLoops) { 2859 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2860 if (!NewLoops.empty()) 2861 U.addSiblingLoops(NewLoops); 2862 2863 // If the current loop remains valid, we should revisit it to catch any 2864 // other unswitch opportunities. Otherwise, we need to mark it as deleted. 2865 if (CurrentLoopValid) 2866 U.revisitCurrentLoop(); 2867 else 2868 U.markLoopAsDeleted(L, LoopName); 2869 }; 2870 2871 Optional<MemorySSAUpdater> MSSAU; 2872 if (AR.MSSA) { 2873 MSSAU = MemorySSAUpdater(AR.MSSA); 2874 if (VerifyMemorySSA) 2875 AR.MSSA->verifyMemorySSA(); 2876 } 2877 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, UnswitchCB, 2878 &AR.SE, MSSAU.hasValue() ? MSSAU.getPointer() : nullptr)) 2879 return PreservedAnalyses::all(); 2880 2881 if (AR.MSSA && VerifyMemorySSA) 2882 AR.MSSA->verifyMemorySSA(); 2883 2884 // Historically this pass has had issues with the dominator tree so verify it 2885 // in asserts builds. 2886 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast)); 2887 2888 auto PA = getLoopPassPreservedAnalyses(); 2889 if (AR.MSSA) 2890 PA.preserve<MemorySSAAnalysis>(); 2891 return PA; 2892 } 2893 2894 namespace { 2895 2896 class SimpleLoopUnswitchLegacyPass : public LoopPass { 2897 bool NonTrivial; 2898 2899 public: 2900 static char ID; // Pass ID, replacement for typeid 2901 2902 explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false) 2903 : LoopPass(ID), NonTrivial(NonTrivial) { 2904 initializeSimpleLoopUnswitchLegacyPassPass( 2905 *PassRegistry::getPassRegistry()); 2906 } 2907 2908 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 2909 2910 void getAnalysisUsage(AnalysisUsage &AU) const override { 2911 AU.addRequired<AssumptionCacheTracker>(); 2912 AU.addRequired<TargetTransformInfoWrapperPass>(); 2913 if (EnableMSSALoopDependency) { 2914 AU.addRequired<MemorySSAWrapperPass>(); 2915 AU.addPreserved<MemorySSAWrapperPass>(); 2916 } 2917 getLoopAnalysisUsage(AU); 2918 } 2919 }; 2920 2921 } // end anonymous namespace 2922 2923 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { 2924 if (skipLoop(L)) 2925 return false; 2926 2927 Function &F = *L->getHeader()->getParent(); 2928 2929 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L 2930 << "\n"); 2931 2932 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2933 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2934 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 2935 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 2936 MemorySSA *MSSA = nullptr; 2937 Optional<MemorySSAUpdater> MSSAU; 2938 if (EnableMSSALoopDependency) { 2939 MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA(); 2940 MSSAU = MemorySSAUpdater(MSSA); 2941 } 2942 2943 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); 2944 auto *SE = SEWP ? &SEWP->getSE() : nullptr; 2945 2946 auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid, 2947 ArrayRef<Loop *> NewLoops) { 2948 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2949 for (auto *NewL : NewLoops) 2950 LPM.addLoop(*NewL); 2951 2952 // If the current loop remains valid, re-add it to the queue. This is 2953 // a little wasteful as we'll finish processing the current loop as well, 2954 // but it is the best we can do in the old PM. 2955 if (CurrentLoopValid) 2956 LPM.addLoop(*L); 2957 else 2958 LPM.markLoopAsDeleted(*L); 2959 }; 2960 2961 if (MSSA && VerifyMemorySSA) 2962 MSSA->verifyMemorySSA(); 2963 2964 bool Changed = unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, UnswitchCB, SE, 2965 MSSAU.hasValue() ? MSSAU.getPointer() : nullptr); 2966 2967 if (MSSA && VerifyMemorySSA) 2968 MSSA->verifyMemorySSA(); 2969 2970 // If anything was unswitched, also clear any cached information about this 2971 // loop. 2972 LPM.deleteSimpleAnalysisLoop(L); 2973 2974 // Historically this pass has had issues with the dominator tree so verify it 2975 // in asserts builds. 2976 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2977 2978 return Changed; 2979 } 2980 2981 char SimpleLoopUnswitchLegacyPass::ID = 0; 2982 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", 2983 "Simple unswitch loops", false, false) 2984 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 2985 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 2986 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 2987 INITIALIZE_PASS_DEPENDENCY(LoopPass) 2988 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 2989 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 2990 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", 2991 "Simple unswitch loops", false, false) 2992 2993 Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) { 2994 return new SimpleLoopUnswitchLegacyPass(NonTrivial); 2995 } 2996