1 //===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 11 #include "llvm/ADT/DenseMap.h" 12 #include "llvm/ADT/STLExtras.h" 13 #include "llvm/ADT/Sequence.h" 14 #include "llvm/ADT/SetVector.h" 15 #include "llvm/ADT/SmallPtrSet.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/ADT/Twine.h" 19 #include "llvm/Analysis/AssumptionCache.h" 20 #include "llvm/Analysis/CFG.h" 21 #include "llvm/Analysis/CodeMetrics.h" 22 #include "llvm/Analysis/LoopAnalysisManager.h" 23 #include "llvm/Analysis/LoopInfo.h" 24 #include "llvm/Analysis/LoopIterator.h" 25 #include "llvm/Analysis/LoopPass.h" 26 #include "llvm/IR/BasicBlock.h" 27 #include "llvm/IR/Constant.h" 28 #include "llvm/IR/Constants.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/Function.h" 31 #include "llvm/IR/InstrTypes.h" 32 #include "llvm/IR/Instruction.h" 33 #include "llvm/IR/Instructions.h" 34 #include "llvm/IR/IntrinsicInst.h" 35 #include "llvm/IR/Use.h" 36 #include "llvm/IR/Value.h" 37 #include "llvm/Pass.h" 38 #include "llvm/Support/Casting.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/GenericDomTree.h" 42 #include "llvm/Support/raw_ostream.h" 43 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 44 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 45 #include "llvm/Transforms/Utils/Cloning.h" 46 #include "llvm/Transforms/Utils/LoopUtils.h" 47 #include "llvm/Transforms/Utils/ValueMapper.h" 48 #include <algorithm> 49 #include <cassert> 50 #include <iterator> 51 #include <numeric> 52 #include <utility> 53 54 #define DEBUG_TYPE "simple-loop-unswitch" 55 56 using namespace llvm; 57 58 STATISTIC(NumBranches, "Number of branches unswitched"); 59 STATISTIC(NumSwitches, "Number of switches unswitched"); 60 STATISTIC(NumTrivial, "Number of unswitches that are trivial"); 61 62 static cl::opt<bool> EnableNonTrivialUnswitch( 63 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden, 64 cl::desc("Forcibly enables non-trivial loop unswitching rather than " 65 "following the configuration passed into the pass.")); 66 67 static cl::opt<int> 68 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, 69 cl::desc("The cost threshold for unswitching a loop.")); 70 71 static void replaceLoopUsesWithConstant(Loop &L, Value &LIC, 72 Constant &Replacement) { 73 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 74 75 // Replace uses of LIC in the loop with the given constant. 76 for (auto UI = LIC.use_begin(), UE = LIC.use_end(); UI != UE;) { 77 // Grab the use and walk past it so we can clobber it in the use list. 78 Use *U = &*UI++; 79 Instruction *UserI = dyn_cast<Instruction>(U->getUser()); 80 if (!UserI || !L.contains(UserI)) 81 continue; 82 83 // Replace this use within the loop body. 84 *U = &Replacement; 85 } 86 } 87 88 /// Update the IDom for a basic block whose predecessor set has changed. 89 /// 90 /// This routine is designed to work when the domtree update is relatively 91 /// localized by leveraging a known common dominator, often a loop header. 92 /// 93 /// FIXME: Should consider hand-rolling a slightly more efficient non-DFS 94 /// approach here as we can do that easily by persisting the candidate IDom's 95 /// dominating set between each predecessor. 96 /// 97 /// FIXME: Longer term, many uses of this can be replaced by an incremental 98 /// domtree update strategy that starts from a known dominating block and 99 /// rebuilds that subtree. 100 static bool updateIDomWithKnownCommonDominator(BasicBlock *BB, 101 BasicBlock *KnownDominatingBB, 102 DominatorTree &DT) { 103 assert(pred_begin(BB) != pred_end(BB) && 104 "This routine does not handle unreachable blocks!"); 105 106 BasicBlock *OrigIDom = DT[BB]->getIDom()->getBlock(); 107 108 BasicBlock *IDom = *pred_begin(BB); 109 assert(DT.dominates(KnownDominatingBB, IDom) && 110 "Bad known dominating block!"); 111 112 // Walk all of the other predecessors finding the nearest common dominator 113 // until all predecessors are covered or we reach the loop header. The loop 114 // header necessarily dominates all loop exit blocks in loop simplified form 115 // so we can early-exit the moment we hit that block. 116 for (auto PI = std::next(pred_begin(BB)), PE = pred_end(BB); 117 PI != PE && IDom != KnownDominatingBB; ++PI) { 118 assert(DT.dominates(KnownDominatingBB, *PI) && 119 "Bad known dominating block!"); 120 IDom = DT.findNearestCommonDominator(IDom, *PI); 121 } 122 123 if (IDom == OrigIDom) 124 return false; 125 126 DT.changeImmediateDominator(BB, IDom); 127 return true; 128 } 129 130 // Note that we don't currently use the IDFCalculator here for two reasons: 131 // 1) It computes dominator tree levels for the entire function on each run 132 // of 'compute'. While this isn't terrible, given that we expect to update 133 // relatively small subtrees of the domtree, it isn't necessarily the right 134 // tradeoff. 135 // 2) The interface doesn't fit this usage well. It doesn't operate in 136 // append-only, and builds several sets that we don't need. 137 // 138 // FIXME: Neither of these issues are a big deal and could be addressed with 139 // some amount of refactoring of IDFCalculator. That would allow us to share 140 // the core logic here (which is solving the same core problem). 141 static void appendDomFrontier(DomTreeNode *Node, 142 SmallSetVector<BasicBlock *, 4> &Worklist, 143 SmallVectorImpl<DomTreeNode *> &DomNodes, 144 SmallPtrSetImpl<BasicBlock *> &DomSet) { 145 assert(DomNodes.empty() && "Must start with no dominator nodes."); 146 assert(DomSet.empty() && "Must start with an empty dominator set."); 147 148 // First flatten this subtree into sequence of nodes by doing a pre-order 149 // walk. 150 DomNodes.push_back(Node); 151 // We intentionally re-evaluate the size as each node can add new children. 152 // Because this is a tree walk, this cannot add any duplicates. 153 for (int i = 0; i < (int)DomNodes.size(); ++i) 154 DomNodes.insert(DomNodes.end(), DomNodes[i]->begin(), DomNodes[i]->end()); 155 156 // Now create a set of the basic blocks so we can quickly test for 157 // dominated successors. We could in theory use the DFS numbers of the 158 // dominator tree for this, but we want this to remain predictably fast 159 // even while we mutate the dominator tree in ways that would invalidate 160 // the DFS numbering. 161 for (DomTreeNode *InnerN : DomNodes) 162 DomSet.insert(InnerN->getBlock()); 163 164 // Now re-walk the nodes, appending every successor of every node that isn't 165 // in the set. Note that we don't append the node itself, even though if it 166 // is a successor it does not strictly dominate itself and thus it would be 167 // part of the dominance frontier. The reason we don't append it is that 168 // the node passed in came *from* the worklist and so it has already been 169 // processed. 170 for (DomTreeNode *InnerN : DomNodes) 171 for (BasicBlock *SuccBB : successors(InnerN->getBlock())) 172 if (!DomSet.count(SuccBB)) 173 Worklist.insert(SuccBB); 174 175 DomNodes.clear(); 176 DomSet.clear(); 177 } 178 179 /// Update the dominator tree after unswitching a particular former exit block. 180 /// 181 /// This handles the full update of the dominator tree after hoisting a block 182 /// that previously was an exit block (or split off of an exit block) up to be 183 /// reached from the new immediate dominator of the preheader. 184 /// 185 /// The common case is simple -- we just move the unswitched block to have an 186 /// immediate dominator of the old preheader. But in complex cases, there may 187 /// be other blocks reachable from the unswitched block that are immediately 188 /// dominated by some node between the unswitched one and the old preheader. 189 /// All of these also need to be hoisted in the dominator tree. We also want to 190 /// minimize queries to the dominator tree because each step of this 191 /// invalidates any DFS numbers that would make queries fast. 192 static void updateDTAfterUnswitch(BasicBlock *UnswitchedBB, BasicBlock *OldPH, 193 DominatorTree &DT) { 194 DomTreeNode *OldPHNode = DT[OldPH]; 195 DomTreeNode *UnswitchedNode = DT[UnswitchedBB]; 196 // If the dominator tree has already been updated for this unswitched node, 197 // we're done. This makes it easier to use this routine if there are multiple 198 // paths to the same unswitched destination. 199 if (UnswitchedNode->getIDom() == OldPHNode) 200 return; 201 202 // First collect the domtree nodes that we are hoisting over. These are the 203 // set of nodes which may have children that need to be hoisted as well. 204 SmallPtrSet<DomTreeNode *, 4> DomChain; 205 for (auto *IDom = UnswitchedNode->getIDom(); IDom != OldPHNode; 206 IDom = IDom->getIDom()) 207 DomChain.insert(IDom); 208 209 // The unswitched block ends up immediately dominated by the old preheader -- 210 // regardless of whether it is the loop exit block or split off of the loop 211 // exit block. 212 DT.changeImmediateDominator(UnswitchedNode, OldPHNode); 213 214 // For everything that moves up the dominator tree, we need to examine the 215 // dominator frontier to see if it additionally should move up the dominator 216 // tree. This lambda appends the dominator frontier for a node on the 217 // worklist. 218 SmallSetVector<BasicBlock *, 4> Worklist; 219 220 // Scratch data structures reused by domfrontier finding. 221 SmallVector<DomTreeNode *, 4> DomNodes; 222 SmallPtrSet<BasicBlock *, 4> DomSet; 223 224 // Append the initial dom frontier nodes. 225 appendDomFrontier(UnswitchedNode, Worklist, DomNodes, DomSet); 226 227 // Walk the worklist. We grow the list in the loop and so must recompute size. 228 for (int i = 0; i < (int)Worklist.size(); ++i) { 229 auto *BB = Worklist[i]; 230 231 DomTreeNode *Node = DT[BB]; 232 assert(!DomChain.count(Node) && 233 "Cannot be dominated by a block you can reach!"); 234 235 // If this block had an immediate dominator somewhere in the chain 236 // we hoisted over, then its position in the domtree needs to move as it is 237 // reachable from a node hoisted over this chain. 238 if (!DomChain.count(Node->getIDom())) 239 continue; 240 241 DT.changeImmediateDominator(Node, OldPHNode); 242 243 // Now add this node's dominator frontier to the worklist as well. 244 appendDomFrontier(Node, Worklist, DomNodes, DomSet); 245 } 246 } 247 248 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial 249 /// incoming values along this edge. 250 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB, 251 BasicBlock &ExitBB) { 252 for (Instruction &I : ExitBB) { 253 auto *PN = dyn_cast<PHINode>(&I); 254 if (!PN) 255 // No more PHIs to check. 256 return true; 257 258 // If the incoming value for this edge isn't loop invariant the unswitch 259 // won't be trivial. 260 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))) 261 return false; 262 } 263 llvm_unreachable("Basic blocks should never be empty!"); 264 } 265 266 /// Rewrite the PHI nodes in an unswitched loop exit basic block. 267 /// 268 /// Requires that the loop exit and unswitched basic block are the same, and 269 /// that the exiting block was a unique predecessor of that block. Rewrites the 270 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial 271 /// PHI nodes from the old preheader that now contains the unswitched 272 /// terminator. 273 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, 274 BasicBlock &OldExitingBB, 275 BasicBlock &OldPH) { 276 for (PHINode &PN : UnswitchedBB.phis()) { 277 // When the loop exit is directly unswitched we just need to update the 278 // incoming basic block. We loop to handle weird cases with repeated 279 // incoming blocks, but expect to typically only have one operand here. 280 for (auto i : seq<int>(0, PN.getNumOperands())) { 281 assert(PN.getIncomingBlock(i) == &OldExitingBB && 282 "Found incoming block different from unique predecessor!"); 283 PN.setIncomingBlock(i, &OldPH); 284 } 285 } 286 } 287 288 /// Rewrite the PHI nodes in the loop exit basic block and the split off 289 /// unswitched block. 290 /// 291 /// Because the exit block remains an exit from the loop, this rewrites the 292 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI 293 /// nodes into the unswitched basic block to select between the value in the 294 /// old preheader and the loop exit. 295 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, 296 BasicBlock &UnswitchedBB, 297 BasicBlock &OldExitingBB, 298 BasicBlock &OldPH) { 299 assert(&ExitBB != &UnswitchedBB && 300 "Must have different loop exit and unswitched blocks!"); 301 Instruction *InsertPt = &*UnswitchedBB.begin(); 302 for (PHINode &PN : ExitBB.phis()) { 303 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2, 304 PN.getName() + ".split", InsertPt); 305 306 // Walk backwards over the old PHI node's inputs to minimize the cost of 307 // removing each one. We have to do this weird loop manually so that we 308 // create the same number of new incoming edges in the new PHI as we expect 309 // each case-based edge to be included in the unswitched switch in some 310 // cases. 311 // FIXME: This is really, really gross. It would be much cleaner if LLVM 312 // allowed us to create a single entry for a predecessor block without 313 // having separate entries for each "edge" even though these edges are 314 // required to produce identical results. 315 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) { 316 if (PN.getIncomingBlock(i) != &OldExitingBB) 317 continue; 318 319 Value *Incoming = PN.removeIncomingValue(i); 320 NewPN->addIncoming(Incoming, &OldPH); 321 } 322 323 // Now replace the old PHI with the new one and wire the old one in as an 324 // input to the new one. 325 PN.replaceAllUsesWith(NewPN); 326 NewPN->addIncoming(&PN, &ExitBB); 327 } 328 } 329 330 /// Unswitch a trivial branch if the condition is loop invariant. 331 /// 332 /// This routine should only be called when loop code leading to the branch has 333 /// been validated as trivial (no side effects). This routine checks if the 334 /// condition is invariant and one of the successors is a loop exit. This 335 /// allows us to unswitch without duplicating the loop, making it trivial. 336 /// 337 /// If this routine fails to unswitch the branch it returns false. 338 /// 339 /// If the branch can be unswitched, this routine splits the preheader and 340 /// hoists the branch above that split. Preserves loop simplified form 341 /// (splitting the exit block as necessary). It simplifies the branch within 342 /// the loop to an unconditional branch but doesn't remove it entirely. Further 343 /// cleanup can be done with some simplify-cfg like pass. 344 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, 345 LoopInfo &LI) { 346 assert(BI.isConditional() && "Can only unswitch a conditional branch!"); 347 DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n"); 348 349 Value *LoopCond = BI.getCondition(); 350 351 // Need a trivial loop condition to unswitch. 352 if (!L.isLoopInvariant(LoopCond)) 353 return false; 354 355 // FIXME: We should compute this once at the start and update it! 356 SmallVector<BasicBlock *, 16> ExitBlocks; 357 L.getExitBlocks(ExitBlocks); 358 SmallPtrSet<BasicBlock *, 16> ExitBlockSet(ExitBlocks.begin(), 359 ExitBlocks.end()); 360 361 // Check to see if a successor of the branch is guaranteed to 362 // exit through a unique exit block without having any 363 // side-effects. If so, determine the value of Cond that causes 364 // it to do this. 365 ConstantInt *CondVal = ConstantInt::getTrue(BI.getContext()); 366 ConstantInt *Replacement = ConstantInt::getFalse(BI.getContext()); 367 int LoopExitSuccIdx = 0; 368 auto *LoopExitBB = BI.getSuccessor(0); 369 if (!ExitBlockSet.count(LoopExitBB)) { 370 std::swap(CondVal, Replacement); 371 LoopExitSuccIdx = 1; 372 LoopExitBB = BI.getSuccessor(1); 373 if (!ExitBlockSet.count(LoopExitBB)) 374 return false; 375 } 376 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); 377 assert(L.contains(ContinueBB) && 378 "Cannot have both successors exit and still be in the loop!"); 379 380 auto *ParentBB = BI.getParent(); 381 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) 382 return false; 383 384 DEBUG(dbgs() << " unswitching trivial branch when: " << CondVal 385 << " == " << LoopCond << "\n"); 386 387 // Split the preheader, so that we know that there is a safe place to insert 388 // the conditional branch. We will change the preheader to have a conditional 389 // branch on LoopCond. 390 BasicBlock *OldPH = L.getLoopPreheader(); 391 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI); 392 393 // Now that we have a place to insert the conditional branch, create a place 394 // to branch to: this is the exit block out of the loop that we are 395 // unswitching. We need to split this if there are other loop predecessors. 396 // Because the loop is in simplified form, *any* other predecessor is enough. 397 BasicBlock *UnswitchedBB; 398 if (BasicBlock *PredBB = LoopExitBB->getUniquePredecessor()) { 399 (void)PredBB; 400 assert(PredBB == BI.getParent() && 401 "A branch's parent isn't a predecessor!"); 402 UnswitchedBB = LoopExitBB; 403 } else { 404 UnswitchedBB = SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI); 405 } 406 407 // Now splice the branch to gate reaching the new preheader and re-point its 408 // successors. 409 OldPH->getInstList().splice(std::prev(OldPH->end()), 410 BI.getParent()->getInstList(), BI); 411 OldPH->getTerminator()->eraseFromParent(); 412 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); 413 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); 414 415 // Create a new unconditional branch that will continue the loop as a new 416 // terminator. 417 BranchInst::Create(ContinueBB, ParentBB); 418 419 // Rewrite the relevant PHI nodes. 420 if (UnswitchedBB == LoopExitBB) 421 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); 422 else 423 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, 424 *ParentBB, *OldPH); 425 426 // Now we need to update the dominator tree. 427 updateDTAfterUnswitch(UnswitchedBB, OldPH, DT); 428 // But if we split something off of the loop exit block then we also removed 429 // one of the predecessors for the loop exit block and may need to update its 430 // idom. 431 if (UnswitchedBB != LoopExitBB) 432 updateIDomWithKnownCommonDominator(LoopExitBB, L.getHeader(), DT); 433 434 // Since this is an i1 condition we can also trivially replace uses of it 435 // within the loop with a constant. 436 replaceLoopUsesWithConstant(L, *LoopCond, *Replacement); 437 438 ++NumTrivial; 439 ++NumBranches; 440 return true; 441 } 442 443 /// Unswitch a trivial switch if the condition is loop invariant. 444 /// 445 /// This routine should only be called when loop code leading to the switch has 446 /// been validated as trivial (no side effects). This routine checks if the 447 /// condition is invariant and that at least one of the successors is a loop 448 /// exit. This allows us to unswitch without duplicating the loop, making it 449 /// trivial. 450 /// 451 /// If this routine fails to unswitch the switch it returns false. 452 /// 453 /// If the switch can be unswitched, this routine splits the preheader and 454 /// copies the switch above that split. If the default case is one of the 455 /// exiting cases, it copies the non-exiting cases and points them at the new 456 /// preheader. If the default case is not exiting, it copies the exiting cases 457 /// and points the default at the preheader. It preserves loop simplified form 458 /// (splitting the exit blocks as necessary). It simplifies the switch within 459 /// the loop by removing now-dead cases. If the default case is one of those 460 /// unswitched, it replaces its destination with a new basic block containing 461 /// only unreachable. Such basic blocks, while technically loop exits, are not 462 /// considered for unswitching so this is a stable transform and the same 463 /// switch will not be revisited. If after unswitching there is only a single 464 /// in-loop successor, the switch is further simplified to an unconditional 465 /// branch. Still more cleanup can be done with some simplify-cfg like pass. 466 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, 467 LoopInfo &LI) { 468 DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n"); 469 Value *LoopCond = SI.getCondition(); 470 471 // If this isn't switching on an invariant condition, we can't unswitch it. 472 if (!L.isLoopInvariant(LoopCond)) 473 return false; 474 475 auto *ParentBB = SI.getParent(); 476 477 // FIXME: We should compute this once at the start and update it! 478 SmallVector<BasicBlock *, 16> ExitBlocks; 479 L.getExitBlocks(ExitBlocks); 480 SmallPtrSet<BasicBlock *, 16> ExitBlockSet(ExitBlocks.begin(), 481 ExitBlocks.end()); 482 483 SmallVector<int, 4> ExitCaseIndices; 484 for (auto Case : SI.cases()) { 485 auto *SuccBB = Case.getCaseSuccessor(); 486 if (ExitBlockSet.count(SuccBB) && 487 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB)) 488 ExitCaseIndices.push_back(Case.getCaseIndex()); 489 } 490 BasicBlock *DefaultExitBB = nullptr; 491 if (ExitBlockSet.count(SI.getDefaultDest()) && 492 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) && 493 !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator())) 494 DefaultExitBB = SI.getDefaultDest(); 495 else if (ExitCaseIndices.empty()) 496 return false; 497 498 DEBUG(dbgs() << " unswitching trivial cases...\n"); 499 500 SmallVector<std::pair<ConstantInt *, BasicBlock *>, 4> ExitCases; 501 ExitCases.reserve(ExitCaseIndices.size()); 502 // We walk the case indices backwards so that we remove the last case first 503 // and don't disrupt the earlier indices. 504 for (unsigned Index : reverse(ExitCaseIndices)) { 505 auto CaseI = SI.case_begin() + Index; 506 // Save the value of this case. 507 ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()}); 508 // Delete the unswitched cases. 509 SI.removeCase(CaseI); 510 } 511 512 // Check if after this all of the remaining cases point at the same 513 // successor. 514 BasicBlock *CommonSuccBB = nullptr; 515 if (SI.getNumCases() > 0 && 516 std::all_of(std::next(SI.case_begin()), SI.case_end(), 517 [&SI](const SwitchInst::CaseHandle &Case) { 518 return Case.getCaseSuccessor() == 519 SI.case_begin()->getCaseSuccessor(); 520 })) 521 CommonSuccBB = SI.case_begin()->getCaseSuccessor(); 522 523 if (DefaultExitBB) { 524 // We can't remove the default edge so replace it with an edge to either 525 // the single common remaining successor (if we have one) or an unreachable 526 // block. 527 if (CommonSuccBB) { 528 SI.setDefaultDest(CommonSuccBB); 529 } else { 530 BasicBlock *UnreachableBB = BasicBlock::Create( 531 ParentBB->getContext(), 532 Twine(ParentBB->getName()) + ".unreachable_default", 533 ParentBB->getParent()); 534 new UnreachableInst(ParentBB->getContext(), UnreachableBB); 535 SI.setDefaultDest(UnreachableBB); 536 DT.addNewBlock(UnreachableBB, ParentBB); 537 } 538 } else { 539 // If we're not unswitching the default, we need it to match any cases to 540 // have a common successor or if we have no cases it is the common 541 // successor. 542 if (SI.getNumCases() == 0) 543 CommonSuccBB = SI.getDefaultDest(); 544 else if (SI.getDefaultDest() != CommonSuccBB) 545 CommonSuccBB = nullptr; 546 } 547 548 // Split the preheader, so that we know that there is a safe place to insert 549 // the switch. 550 BasicBlock *OldPH = L.getLoopPreheader(); 551 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI); 552 OldPH->getTerminator()->eraseFromParent(); 553 554 // Now add the unswitched switch. 555 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); 556 557 // Rewrite the IR for the unswitched basic blocks. This requires two steps. 558 // First, we split any exit blocks with remaining in-loop predecessors. Then 559 // we update the PHIs in one of two ways depending on if there was a split. 560 // We walk in reverse so that we split in the same order as the cases 561 // appeared. This is purely for convenience of reading the resulting IR, but 562 // it doesn't cost anything really. 563 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; 564 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; 565 // Handle the default exit if necessary. 566 // FIXME: It'd be great if we could merge this with the loop below but LLVM's 567 // ranges aren't quite powerful enough yet. 568 if (DefaultExitBB) { 569 if (pred_empty(DefaultExitBB)) { 570 UnswitchedExitBBs.insert(DefaultExitBB); 571 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); 572 } else { 573 auto *SplitBB = 574 SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI); 575 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB, 576 *ParentBB, *OldPH); 577 updateIDomWithKnownCommonDominator(DefaultExitBB, L.getHeader(), DT); 578 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; 579 } 580 } 581 // Note that we must use a reference in the for loop so that we update the 582 // container. 583 for (auto &CasePair : reverse(ExitCases)) { 584 // Grab a reference to the exit block in the pair so that we can update it. 585 BasicBlock *ExitBB = CasePair.second; 586 587 // If this case is the last edge into the exit block, we can simply reuse it 588 // as it will no longer be a loop exit. No mapping necessary. 589 if (pred_empty(ExitBB)) { 590 // Only rewrite once. 591 if (UnswitchedExitBBs.insert(ExitBB).second) 592 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); 593 continue; 594 } 595 596 // Otherwise we need to split the exit block so that we retain an exit 597 // block from the loop and a target for the unswitched condition. 598 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; 599 if (!SplitExitBB) { 600 // If this is the first time we see this, do the split and remember it. 601 SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI); 602 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB, 603 *ParentBB, *OldPH); 604 updateIDomWithKnownCommonDominator(ExitBB, L.getHeader(), DT); 605 } 606 // Update the case pair to point to the split block. 607 CasePair.second = SplitExitBB; 608 } 609 610 // Now add the unswitched cases. We do this in reverse order as we built them 611 // in reverse order. 612 for (auto CasePair : reverse(ExitCases)) { 613 ConstantInt *CaseVal = CasePair.first; 614 BasicBlock *UnswitchedBB = CasePair.second; 615 616 NewSI->addCase(CaseVal, UnswitchedBB); 617 updateDTAfterUnswitch(UnswitchedBB, OldPH, DT); 618 } 619 620 // If the default was unswitched, re-point it and add explicit cases for 621 // entering the loop. 622 if (DefaultExitBB) { 623 NewSI->setDefaultDest(DefaultExitBB); 624 updateDTAfterUnswitch(DefaultExitBB, OldPH, DT); 625 626 // We removed all the exit cases, so we just copy the cases to the 627 // unswitched switch. 628 for (auto Case : SI.cases()) 629 NewSI->addCase(Case.getCaseValue(), NewPH); 630 } 631 632 // If we ended up with a common successor for every path through the switch 633 // after unswitching, rewrite it to an unconditional branch to make it easy 634 // to recognize. Otherwise we potentially have to recognize the default case 635 // pointing at unreachable and other complexity. 636 if (CommonSuccBB) { 637 BasicBlock *BB = SI.getParent(); 638 SI.eraseFromParent(); 639 BranchInst::Create(CommonSuccBB, BB); 640 } 641 642 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 643 ++NumTrivial; 644 ++NumSwitches; 645 return true; 646 } 647 648 /// This routine scans the loop to find a branch or switch which occurs before 649 /// any side effects occur. These can potentially be unswitched without 650 /// duplicating the loop. If a branch or switch is successfully unswitched the 651 /// scanning continues to see if subsequent branches or switches have become 652 /// trivial. Once all trivial candidates have been unswitched, this routine 653 /// returns. 654 /// 655 /// The return value indicates whether anything was unswitched (and therefore 656 /// changed). 657 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, 658 LoopInfo &LI) { 659 bool Changed = false; 660 661 // If loop header has only one reachable successor we should keep looking for 662 // trivial condition candidates in the successor as well. An alternative is 663 // to constant fold conditions and merge successors into loop header (then we 664 // only need to check header's terminator). The reason for not doing this in 665 // LoopUnswitch pass is that it could potentially break LoopPassManager's 666 // invariants. Folding dead branches could either eliminate the current loop 667 // or make other loops unreachable. LCSSA form might also not be preserved 668 // after deleting branches. The following code keeps traversing loop header's 669 // successors until it finds the trivial condition candidate (condition that 670 // is not a constant). Since unswitching generates branches with constant 671 // conditions, this scenario could be very common in practice. 672 BasicBlock *CurrentBB = L.getHeader(); 673 SmallPtrSet<BasicBlock *, 8> Visited; 674 Visited.insert(CurrentBB); 675 do { 676 // Check if there are any side-effecting instructions (e.g. stores, calls, 677 // volatile loads) in the part of the loop that the code *would* execute 678 // without unswitching. 679 if (llvm::any_of(*CurrentBB, 680 [](Instruction &I) { return I.mayHaveSideEffects(); })) 681 return Changed; 682 683 TerminatorInst *CurrentTerm = CurrentBB->getTerminator(); 684 685 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) { 686 // Don't bother trying to unswitch past a switch with a constant 687 // condition. This should be removed prior to running this pass by 688 // simplify-cfg. 689 if (isa<Constant>(SI->getCondition())) 690 return Changed; 691 692 if (!unswitchTrivialSwitch(L, *SI, DT, LI)) 693 // Coludn't unswitch this one so we're done. 694 return Changed; 695 696 // Mark that we managed to unswitch something. 697 Changed = true; 698 699 // If unswitching turned the terminator into an unconditional branch then 700 // we can continue. The unswitching logic specifically works to fold any 701 // cases it can into an unconditional branch to make it easier to 702 // recognize here. 703 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); 704 if (!BI || BI->isConditional()) 705 return Changed; 706 707 CurrentBB = BI->getSuccessor(0); 708 continue; 709 } 710 711 auto *BI = dyn_cast<BranchInst>(CurrentTerm); 712 if (!BI) 713 // We do not understand other terminator instructions. 714 return Changed; 715 716 // Don't bother trying to unswitch past an unconditional branch or a branch 717 // with a constant value. These should be removed by simplify-cfg prior to 718 // running this pass. 719 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 720 return Changed; 721 722 // Found a trivial condition candidate: non-foldable conditional branch. If 723 // we fail to unswitch this, we can't do anything else that is trivial. 724 if (!unswitchTrivialBranch(L, *BI, DT, LI)) 725 return Changed; 726 727 // Mark that we managed to unswitch something. 728 Changed = true; 729 730 // We unswitched the branch. This should always leave us with an 731 // unconditional branch that we can follow now. 732 BI = cast<BranchInst>(CurrentBB->getTerminator()); 733 assert(!BI->isConditional() && 734 "Cannot form a conditional branch by unswitching1"); 735 CurrentBB = BI->getSuccessor(0); 736 737 // When continuing, if we exit the loop or reach a previous visited block, 738 // then we can not reach any trivial condition candidates (unfoldable 739 // branch instructions or switch instructions) and no unswitch can happen. 740 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second); 741 742 return Changed; 743 } 744 745 /// Build the cloned blocks for an unswitched copy of the given loop. 746 /// 747 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and 748 /// after the split block (`SplitBB`) that will be used to select between the 749 /// cloned and original loop. 750 /// 751 /// This routine handles cloning all of the necessary loop blocks and exit 752 /// blocks including rewriting their instructions and the relevant PHI nodes. 753 /// It skips loop and exit blocks that are not necessary based on the provided 754 /// set. It also correctly creates the unconditional branch in the cloned 755 /// unswitched parent block to only point at the unswitched successor. 756 /// 757 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit 758 /// block splitting is correctly reflected in `LoopInfo`, essentially all of 759 /// the cloned blocks (and their loops) are left without full `LoopInfo` 760 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned 761 /// blocks to them but doesn't create the cloned `DominatorTree` structure and 762 /// instead the caller must recompute an accurate DT. It *does* correctly 763 /// update the `AssumptionCache` provided in `AC`. 764 static BasicBlock *buildClonedLoopBlocks( 765 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, 766 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB, 767 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, 768 const SmallPtrSetImpl<BasicBlock *> &SkippedLoopAndExitBlocks, 769 ValueToValueMapTy &VMap, 770 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC, 771 DominatorTree &DT, LoopInfo &LI) { 772 SmallVector<BasicBlock *, 4> NewBlocks; 773 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size()); 774 775 // We will need to clone a bunch of blocks, wrap up the clone operation in 776 // a helper. 777 auto CloneBlock = [&](BasicBlock *OldBB) { 778 // Clone the basic block and insert it before the new preheader. 779 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent()); 780 NewBB->moveBefore(LoopPH); 781 782 // Record this block and the mapping. 783 NewBlocks.push_back(NewBB); 784 VMap[OldBB] = NewBB; 785 786 return NewBB; 787 }; 788 789 // First, clone the preheader. 790 auto *ClonedPH = CloneBlock(LoopPH); 791 792 // Then clone all the loop blocks, skipping the ones that aren't necessary. 793 for (auto *LoopBB : L.blocks()) 794 if (!SkippedLoopAndExitBlocks.count(LoopBB)) 795 CloneBlock(LoopBB); 796 797 // Split all the loop exit edges so that when we clone the exit blocks, if 798 // any of the exit blocks are *also* a preheader for some other loop, we 799 // don't create multiple predecessors entering the loop header. 800 for (auto *ExitBB : ExitBlocks) { 801 if (SkippedLoopAndExitBlocks.count(ExitBB)) 802 continue; 803 804 // When we are going to clone an exit, we don't need to clone all the 805 // instructions in the exit block and we want to ensure we have an easy 806 // place to merge the CFG, so split the exit first. This is always safe to 807 // do because there cannot be any non-loop predecessors of a loop exit in 808 // loop simplified form. 809 auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI); 810 811 // Rearrange the names to make it easier to write test cases by having the 812 // exit block carry the suffix rather than the merge block carrying the 813 // suffix. 814 MergeBB->takeName(ExitBB); 815 ExitBB->setName(Twine(MergeBB->getName()) + ".split"); 816 817 // Now clone the original exit block. 818 auto *ClonedExitBB = CloneBlock(ExitBB); 819 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 && 820 "Exit block should have been split to have one successor!"); 821 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB && 822 "Cloned exit block has the wrong successor!"); 823 824 // Remap any cloned instructions and create a merge phi node for them. 825 for (auto ZippedInsts : llvm::zip_first( 826 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())), 827 llvm::make_range(ClonedExitBB->begin(), 828 std::prev(ClonedExitBB->end())))) { 829 Instruction &I = std::get<0>(ZippedInsts); 830 Instruction &ClonedI = std::get<1>(ZippedInsts); 831 832 // The only instructions in the exit block should be PHI nodes and 833 // potentially a landing pad. 834 assert( 835 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) && 836 "Bad instruction in exit block!"); 837 // We should have a value map between the instruction and its clone. 838 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!"); 839 840 auto *MergePN = 841 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi", 842 &*MergeBB->getFirstInsertionPt()); 843 I.replaceAllUsesWith(MergePN); 844 MergePN->addIncoming(&I, ExitBB); 845 MergePN->addIncoming(&ClonedI, ClonedExitBB); 846 } 847 } 848 849 // Rewrite the instructions in the cloned blocks to refer to the instructions 850 // in the cloned blocks. We have to do this as a second pass so that we have 851 // everything available. Also, we have inserted new instructions which may 852 // include assume intrinsics, so we update the assumption cache while 853 // processing this. 854 for (auto *ClonedBB : NewBlocks) 855 for (Instruction &I : *ClonedBB) { 856 RemapInstruction(&I, VMap, 857 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 858 if (auto *II = dyn_cast<IntrinsicInst>(&I)) 859 if (II->getIntrinsicID() == Intrinsic::assume) 860 AC.registerAssumption(II); 861 } 862 863 // Remove the cloned parent as a predecessor of the cloned continue successor 864 // if we did in fact clone it. 865 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB)); 866 if (auto *ClonedContinueSuccBB = 867 cast_or_null<BasicBlock>(VMap.lookup(ContinueSuccBB))) 868 ClonedContinueSuccBB->removePredecessor(ClonedParentBB, 869 /*DontDeleteUselessPHIs*/ true); 870 // Replace the cloned branch with an unconditional branch to the cloned 871 // unswitched successor. 872 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB)); 873 ClonedParentBB->getTerminator()->eraseFromParent(); 874 BranchInst::Create(ClonedSuccBB, ClonedParentBB); 875 876 // Update any PHI nodes in the cloned successors of the skipped blocks to not 877 // have spurious incoming values. 878 for (auto *LoopBB : L.blocks()) 879 if (SkippedLoopAndExitBlocks.count(LoopBB)) 880 for (auto *SuccBB : successors(LoopBB)) 881 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB))) 882 for (PHINode &PN : ClonedSuccBB->phis()) 883 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false); 884 885 // Record the domtree updates for the new blocks. 886 SmallPtrSet<BasicBlock *, 4> SuccSet; 887 for (auto *ClonedBB : NewBlocks) { 888 for (auto *SuccBB : successors(ClonedBB)) 889 if (SuccSet.insert(SuccBB).second) 890 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB}); 891 SuccSet.clear(); 892 } 893 894 return ClonedPH; 895 } 896 897 /// Recursively clone the specified loop and all of its children. 898 /// 899 /// The target parent loop for the clone should be provided, or can be null if 900 /// the clone is a top-level loop. While cloning, all the blocks are mapped 901 /// with the provided value map. The entire original loop must be present in 902 /// the value map. The cloned loop is returned. 903 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, 904 const ValueToValueMapTy &VMap, LoopInfo &LI) { 905 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) { 906 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!"); 907 ClonedL.reserveBlocks(OrigL.getNumBlocks()); 908 for (auto *BB : OrigL.blocks()) { 909 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB)); 910 ClonedL.addBlockEntry(ClonedBB); 911 if (LI.getLoopFor(BB) == &OrigL) 912 LI.changeLoopFor(ClonedBB, &ClonedL); 913 } 914 }; 915 916 // We specially handle the first loop because it may get cloned into 917 // a different parent and because we most commonly are cloning leaf loops. 918 Loop *ClonedRootL = LI.AllocateLoop(); 919 if (RootParentL) 920 RootParentL->addChildLoop(ClonedRootL); 921 else 922 LI.addTopLevelLoop(ClonedRootL); 923 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL); 924 925 if (OrigRootL.empty()) 926 return ClonedRootL; 927 928 // If we have a nest, we can quickly clone the entire loop nest using an 929 // iterative approach because it is a tree. We keep the cloned parent in the 930 // data structure to avoid repeatedly querying through a map to find it. 931 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone; 932 // Build up the loops to clone in reverse order as we'll clone them from the 933 // back. 934 for (Loop *ChildL : llvm::reverse(OrigRootL)) 935 LoopsToClone.push_back({ClonedRootL, ChildL}); 936 do { 937 Loop *ClonedParentL, *L; 938 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val(); 939 Loop *ClonedL = LI.AllocateLoop(); 940 ClonedParentL->addChildLoop(ClonedL); 941 AddClonedBlocksToLoop(*L, *ClonedL); 942 for (Loop *ChildL : llvm::reverse(*L)) 943 LoopsToClone.push_back({ClonedL, ChildL}); 944 } while (!LoopsToClone.empty()); 945 946 return ClonedRootL; 947 } 948 949 /// Build the cloned loops of an original loop from unswitching. 950 /// 951 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial 952 /// operation. We need to re-verify that there even is a loop (as the backedge 953 /// may not have been cloned), and even if there are remaining backedges the 954 /// backedge set may be different. However, we know that each child loop is 955 /// undisturbed, we only need to find where to place each child loop within 956 /// either any parent loop or within a cloned version of the original loop. 957 /// 958 /// Because child loops may end up cloned outside of any cloned version of the 959 /// original loop, multiple cloned sibling loops may be created. All of them 960 /// are returned so that the newly introduced loop nest roots can be 961 /// identified. 962 static Loop *buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks, 963 const ValueToValueMapTy &VMap, LoopInfo &LI, 964 SmallVectorImpl<Loop *> &NonChildClonedLoops) { 965 Loop *ClonedL = nullptr; 966 967 auto *OrigPH = OrigL.getLoopPreheader(); 968 auto *OrigHeader = OrigL.getHeader(); 969 970 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH)); 971 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader)); 972 973 // We need to know the loops of the cloned exit blocks to even compute the 974 // accurate parent loop. If we only clone exits to some parent of the 975 // original parent, we want to clone into that outer loop. We also keep track 976 // of the loops that our cloned exit blocks participate in. 977 Loop *ParentL = nullptr; 978 SmallVector<BasicBlock *, 4> ClonedExitsInLoops; 979 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap; 980 ClonedExitsInLoops.reserve(ExitBlocks.size()); 981 for (auto *ExitBB : ExitBlocks) 982 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB))) 983 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 984 ExitLoopMap[ClonedExitBB] = ExitL; 985 ClonedExitsInLoops.push_back(ClonedExitBB); 986 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 987 ParentL = ExitL; 988 } 989 assert((!ParentL || ParentL == OrigL.getParentLoop() || 990 ParentL->contains(OrigL.getParentLoop())) && 991 "The computed parent loop should always contain (or be) the parent of " 992 "the original loop."); 993 994 // We build the set of blocks dominated by the cloned header from the set of 995 // cloned blocks out of the original loop. While not all of these will 996 // necessarily be in the cloned loop, it is enough to establish that they 997 // aren't in unreachable cycles, etc. 998 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks; 999 for (auto *BB : OrigL.blocks()) 1000 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB))) 1001 ClonedLoopBlocks.insert(ClonedBB); 1002 1003 // Rebuild the set of blocks that will end up in the cloned loop. We may have 1004 // skipped cloning some region of this loop which can in turn skip some of 1005 // the backedges so we have to rebuild the blocks in the loop based on the 1006 // backedges that remain after cloning. 1007 SmallVector<BasicBlock *, 16> Worklist; 1008 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop; 1009 for (auto *Pred : predecessors(ClonedHeader)) { 1010 // The only possible non-loop header predecessor is the preheader because 1011 // we know we cloned the loop in simplified form. 1012 if (Pred == ClonedPH) 1013 continue; 1014 1015 // Because the loop was in simplified form, the only non-loop predecessor 1016 // should be the preheader. 1017 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop " 1018 "header other than the preheader " 1019 "that is not part of the loop!"); 1020 1021 // Insert this block into the loop set and on the first visit (and if it 1022 // isn't the header we're currently walking) put it into the worklist to 1023 // recurse through. 1024 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader) 1025 Worklist.push_back(Pred); 1026 } 1027 1028 // If we had any backedges then there *is* a cloned loop. Put the header into 1029 // the loop set and then walk the worklist backwards to find all the blocks 1030 // that remain within the loop after cloning. 1031 if (!BlocksInClonedLoop.empty()) { 1032 BlocksInClonedLoop.insert(ClonedHeader); 1033 1034 while (!Worklist.empty()) { 1035 BasicBlock *BB = Worklist.pop_back_val(); 1036 assert(BlocksInClonedLoop.count(BB) && 1037 "Didn't put block into the loop set!"); 1038 1039 // Insert any predecessors that are in the possible set into the cloned 1040 // set, and if the insert is successful, add them to the worklist. Note 1041 // that we filter on the blocks that are definitely reachable via the 1042 // backedge to the loop header so we may prune out dead code within the 1043 // cloned loop. 1044 for (auto *Pred : predecessors(BB)) 1045 if (ClonedLoopBlocks.count(Pred) && 1046 BlocksInClonedLoop.insert(Pred).second) 1047 Worklist.push_back(Pred); 1048 } 1049 1050 ClonedL = LI.AllocateLoop(); 1051 if (ParentL) { 1052 ParentL->addBasicBlockToLoop(ClonedPH, LI); 1053 ParentL->addChildLoop(ClonedL); 1054 } else { 1055 LI.addTopLevelLoop(ClonedL); 1056 } 1057 1058 ClonedL->reserveBlocks(BlocksInClonedLoop.size()); 1059 // We don't want to just add the cloned loop blocks based on how we 1060 // discovered them. The original order of blocks was carefully built in 1061 // a way that doesn't rely on predecessor ordering. Rather than re-invent 1062 // that logic, we just re-walk the original blocks (and those of the child 1063 // loops) and filter them as we add them into the cloned loop. 1064 for (auto *BB : OrigL.blocks()) { 1065 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)); 1066 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB)) 1067 continue; 1068 1069 // Directly add the blocks that are only in this loop. 1070 if (LI.getLoopFor(BB) == &OrigL) { 1071 ClonedL->addBasicBlockToLoop(ClonedBB, LI); 1072 continue; 1073 } 1074 1075 // We want to manually add it to this loop and parents. 1076 // Registering it with LoopInfo will happen when we clone the top 1077 // loop for this block. 1078 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop()) 1079 PL->addBlockEntry(ClonedBB); 1080 } 1081 1082 // Now add each child loop whose header remains within the cloned loop. All 1083 // of the blocks within the loop must satisfy the same constraints as the 1084 // header so once we pass the header checks we can just clone the entire 1085 // child loop nest. 1086 for (Loop *ChildL : OrigL) { 1087 auto *ClonedChildHeader = 1088 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1089 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader)) 1090 continue; 1091 1092 #ifndef NDEBUG 1093 // We should never have a cloned child loop header but fail to have 1094 // all of the blocks for that child loop. 1095 for (auto *ChildLoopBB : ChildL->blocks()) 1096 assert(BlocksInClonedLoop.count( 1097 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) && 1098 "Child cloned loop has a header within the cloned outer " 1099 "loop but not all of its blocks!"); 1100 #endif 1101 1102 cloneLoopNest(*ChildL, ClonedL, VMap, LI); 1103 } 1104 } 1105 1106 // Now that we've handled all the components of the original loop that were 1107 // cloned into a new loop, we still need to handle anything from the original 1108 // loop that wasn't in a cloned loop. 1109 1110 // Figure out what blocks are left to place within any loop nest containing 1111 // the unswitched loop. If we never formed a loop, the cloned PH is one of 1112 // them. 1113 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet; 1114 if (BlocksInClonedLoop.empty()) 1115 UnloopedBlockSet.insert(ClonedPH); 1116 for (auto *ClonedBB : ClonedLoopBlocks) 1117 if (!BlocksInClonedLoop.count(ClonedBB)) 1118 UnloopedBlockSet.insert(ClonedBB); 1119 1120 // Copy the cloned exits and sort them in ascending loop depth, we'll work 1121 // backwards across these to process them inside out. The order shouldn't 1122 // matter as we're just trying to build up the map from inside-out; we use 1123 // the map in a more stably ordered way below. 1124 auto OrderedClonedExitsInLoops = ClonedExitsInLoops; 1125 llvm::sort(OrderedClonedExitsInLoops.begin(), 1126 OrderedClonedExitsInLoops.end(), 1127 [&](BasicBlock *LHS, BasicBlock *RHS) { 1128 return ExitLoopMap.lookup(LHS)->getLoopDepth() < 1129 ExitLoopMap.lookup(RHS)->getLoopDepth(); 1130 }); 1131 1132 // Populate the existing ExitLoopMap with everything reachable from each 1133 // exit, starting from the inner most exit. 1134 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) { 1135 assert(Worklist.empty() && "Didn't clear worklist!"); 1136 1137 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val(); 1138 Loop *ExitL = ExitLoopMap.lookup(ExitBB); 1139 1140 // Walk the CFG back until we hit the cloned PH adding everything reachable 1141 // and in the unlooped set to this exit block's loop. 1142 Worklist.push_back(ExitBB); 1143 do { 1144 BasicBlock *BB = Worklist.pop_back_val(); 1145 // We can stop recursing at the cloned preheader (if we get there). 1146 if (BB == ClonedPH) 1147 continue; 1148 1149 for (BasicBlock *PredBB : predecessors(BB)) { 1150 // If this pred has already been moved to our set or is part of some 1151 // (inner) loop, no update needed. 1152 if (!UnloopedBlockSet.erase(PredBB)) { 1153 assert( 1154 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) && 1155 "Predecessor not mapped to a loop!"); 1156 continue; 1157 } 1158 1159 // We just insert into the loop set here. We'll add these blocks to the 1160 // exit loop after we build up the set in an order that doesn't rely on 1161 // predecessor order (which in turn relies on use list order). 1162 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second; 1163 (void)Inserted; 1164 assert(Inserted && "Should only visit an unlooped block once!"); 1165 1166 // And recurse through to its predecessors. 1167 Worklist.push_back(PredBB); 1168 } 1169 } while (!Worklist.empty()); 1170 } 1171 1172 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned 1173 // blocks to their outer loops, walk the cloned blocks and the cloned exits 1174 // in their original order adding them to the correct loop. 1175 1176 // We need a stable insertion order. We use the order of the original loop 1177 // order and map into the correct parent loop. 1178 for (auto *BB : llvm::concat<BasicBlock *const>( 1179 makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops)) 1180 if (Loop *OuterL = ExitLoopMap.lookup(BB)) 1181 OuterL->addBasicBlockToLoop(BB, LI); 1182 1183 #ifndef NDEBUG 1184 for (auto &BBAndL : ExitLoopMap) { 1185 auto *BB = BBAndL.first; 1186 auto *OuterL = BBAndL.second; 1187 assert(LI.getLoopFor(BB) == OuterL && 1188 "Failed to put all blocks into outer loops!"); 1189 } 1190 #endif 1191 1192 // Now that all the blocks are placed into the correct containing loop in the 1193 // absence of child loops, find all the potentially cloned child loops and 1194 // clone them into whatever outer loop we placed their header into. 1195 for (Loop *ChildL : OrigL) { 1196 auto *ClonedChildHeader = 1197 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); 1198 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader)) 1199 continue; 1200 1201 #ifndef NDEBUG 1202 for (auto *ChildLoopBB : ChildL->blocks()) 1203 assert(VMap.count(ChildLoopBB) && 1204 "Cloned a child loop header but not all of that loops blocks!"); 1205 #endif 1206 1207 NonChildClonedLoops.push_back(cloneLoopNest( 1208 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI)); 1209 } 1210 1211 // Return the main cloned loop if any. 1212 return ClonedL; 1213 } 1214 1215 static void 1216 deleteDeadBlocksFromLoop(Loop &L, 1217 const SmallVectorImpl<BasicBlock *> &DeadBlocks, 1218 SmallVectorImpl<BasicBlock *> &ExitBlocks, 1219 DominatorTree &DT, LoopInfo &LI) { 1220 SmallPtrSet<BasicBlock *, 16> DeadBlockSet(DeadBlocks.begin(), 1221 DeadBlocks.end()); 1222 1223 // Filter out the dead blocks from the exit blocks list so that it can be 1224 // used in the caller. 1225 llvm::erase_if(ExitBlocks, 1226 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1227 1228 // Remove these blocks from their successors. 1229 for (auto *BB : DeadBlocks) 1230 for (BasicBlock *SuccBB : successors(BB)) 1231 SuccBB->removePredecessor(BB, /*DontDeleteUselessPHIs*/ true); 1232 1233 // Walk from this loop up through its parents removing all of the dead blocks. 1234 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) { 1235 for (auto *BB : DeadBlocks) 1236 ParentL->getBlocksSet().erase(BB); 1237 llvm::erase_if(ParentL->getBlocksVector(), 1238 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); 1239 } 1240 1241 // Now delete the dead child loops. This raw delete will clear them 1242 // recursively. 1243 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) { 1244 if (!DeadBlockSet.count(ChildL->getHeader())) 1245 return false; 1246 1247 assert(llvm::all_of(ChildL->blocks(), 1248 [&](BasicBlock *ChildBB) { 1249 return DeadBlockSet.count(ChildBB); 1250 }) && 1251 "If the child loop header is dead all blocks in the child loop must " 1252 "be dead as well!"); 1253 LI.destroy(ChildL); 1254 return true; 1255 }); 1256 1257 // Remove the loop mappings for the dead blocks and drop all the references 1258 // from these blocks to others to handle cyclic references as we start 1259 // deleting the blocks themselves. 1260 for (auto *BB : DeadBlocks) { 1261 // Check that the dominator tree has already been updated. 1262 assert(!DT.getNode(BB) && "Should already have cleared domtree!"); 1263 LI.changeLoopFor(BB, nullptr); 1264 BB->dropAllReferences(); 1265 } 1266 1267 // Actually delete the blocks now that they've been fully unhooked from the 1268 // IR. 1269 for (auto *BB : DeadBlocks) 1270 BB->eraseFromParent(); 1271 } 1272 1273 /// Recompute the set of blocks in a loop after unswitching. 1274 /// 1275 /// This walks from the original headers predecessors to rebuild the loop. We 1276 /// take advantage of the fact that new blocks can't have been added, and so we 1277 /// filter by the original loop's blocks. This also handles potentially 1278 /// unreachable code that we don't want to explore but might be found examining 1279 /// the predecessors of the header. 1280 /// 1281 /// If the original loop is no longer a loop, this will return an empty set. If 1282 /// it remains a loop, all the blocks within it will be added to the set 1283 /// (including those blocks in inner loops). 1284 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L, 1285 LoopInfo &LI) { 1286 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet; 1287 1288 auto *PH = L.getLoopPreheader(); 1289 auto *Header = L.getHeader(); 1290 1291 // A worklist to use while walking backwards from the header. 1292 SmallVector<BasicBlock *, 16> Worklist; 1293 1294 // First walk the predecessors of the header to find the backedges. This will 1295 // form the basis of our walk. 1296 for (auto *Pred : predecessors(Header)) { 1297 // Skip the preheader. 1298 if (Pred == PH) 1299 continue; 1300 1301 // Because the loop was in simplified form, the only non-loop predecessor 1302 // is the preheader. 1303 assert(L.contains(Pred) && "Found a predecessor of the loop header other " 1304 "than the preheader that is not part of the " 1305 "loop!"); 1306 1307 // Insert this block into the loop set and on the first visit and, if it 1308 // isn't the header we're currently walking, put it into the worklist to 1309 // recurse through. 1310 if (LoopBlockSet.insert(Pred).second && Pred != Header) 1311 Worklist.push_back(Pred); 1312 } 1313 1314 // If no backedges were found, we're done. 1315 if (LoopBlockSet.empty()) 1316 return LoopBlockSet; 1317 1318 // We found backedges, recurse through them to identify the loop blocks. 1319 while (!Worklist.empty()) { 1320 BasicBlock *BB = Worklist.pop_back_val(); 1321 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!"); 1322 1323 // No need to walk past the header. 1324 if (BB == Header) 1325 continue; 1326 1327 // Because we know the inner loop structure remains valid we can use the 1328 // loop structure to jump immediately across the entire nested loop. 1329 // Further, because it is in loop simplified form, we can directly jump 1330 // to its preheader afterward. 1331 if (Loop *InnerL = LI.getLoopFor(BB)) 1332 if (InnerL != &L) { 1333 assert(L.contains(InnerL) && 1334 "Should not reach a loop *outside* this loop!"); 1335 // The preheader is the only possible predecessor of the loop so 1336 // insert it into the set and check whether it was already handled. 1337 auto *InnerPH = InnerL->getLoopPreheader(); 1338 assert(L.contains(InnerPH) && "Cannot contain an inner loop block " 1339 "but not contain the inner loop " 1340 "preheader!"); 1341 if (!LoopBlockSet.insert(InnerPH).second) 1342 // The only way to reach the preheader is through the loop body 1343 // itself so if it has been visited the loop is already handled. 1344 continue; 1345 1346 // Insert all of the blocks (other than those already present) into 1347 // the loop set. We expect at least the block that led us to find the 1348 // inner loop to be in the block set, but we may also have other loop 1349 // blocks if they were already enqueued as predecessors of some other 1350 // outer loop block. 1351 for (auto *InnerBB : InnerL->blocks()) { 1352 if (InnerBB == BB) { 1353 assert(LoopBlockSet.count(InnerBB) && 1354 "Block should already be in the set!"); 1355 continue; 1356 } 1357 1358 LoopBlockSet.insert(InnerBB); 1359 } 1360 1361 // Add the preheader to the worklist so we will continue past the 1362 // loop body. 1363 Worklist.push_back(InnerPH); 1364 continue; 1365 } 1366 1367 // Insert any predecessors that were in the original loop into the new 1368 // set, and if the insert is successful, add them to the worklist. 1369 for (auto *Pred : predecessors(BB)) 1370 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second) 1371 Worklist.push_back(Pred); 1372 } 1373 1374 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!"); 1375 1376 // We've found all the blocks participating in the loop, return our completed 1377 // set. 1378 return LoopBlockSet; 1379 } 1380 1381 /// Rebuild a loop after unswitching removes some subset of blocks and edges. 1382 /// 1383 /// The removal may have removed some child loops entirely but cannot have 1384 /// disturbed any remaining child loops. However, they may need to be hoisted 1385 /// to the parent loop (or to be top-level loops). The original loop may be 1386 /// completely removed. 1387 /// 1388 /// The sibling loops resulting from this update are returned. If the original 1389 /// loop remains a valid loop, it will be the first entry in this list with all 1390 /// of the newly sibling loops following it. 1391 /// 1392 /// Returns true if the loop remains a loop after unswitching, and false if it 1393 /// is no longer a loop after unswitching (and should not continue to be 1394 /// referenced). 1395 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, 1396 LoopInfo &LI, 1397 SmallVectorImpl<Loop *> &HoistedLoops) { 1398 auto *PH = L.getLoopPreheader(); 1399 1400 // Compute the actual parent loop from the exit blocks. Because we may have 1401 // pruned some exits the loop may be different from the original parent. 1402 Loop *ParentL = nullptr; 1403 SmallVector<Loop *, 4> ExitLoops; 1404 SmallVector<BasicBlock *, 4> ExitsInLoops; 1405 ExitsInLoops.reserve(ExitBlocks.size()); 1406 for (auto *ExitBB : ExitBlocks) 1407 if (Loop *ExitL = LI.getLoopFor(ExitBB)) { 1408 ExitLoops.push_back(ExitL); 1409 ExitsInLoops.push_back(ExitBB); 1410 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) 1411 ParentL = ExitL; 1412 } 1413 1414 // Recompute the blocks participating in this loop. This may be empty if it 1415 // is no longer a loop. 1416 auto LoopBlockSet = recomputeLoopBlockSet(L, LI); 1417 1418 // If we still have a loop, we need to re-set the loop's parent as the exit 1419 // block set changing may have moved it within the loop nest. Note that this 1420 // can only happen when this loop has a parent as it can only hoist the loop 1421 // *up* the nest. 1422 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) { 1423 // Remove this loop's (original) blocks from all of the intervening loops. 1424 for (Loop *IL = L.getParentLoop(); IL != ParentL; 1425 IL = IL->getParentLoop()) { 1426 IL->getBlocksSet().erase(PH); 1427 for (auto *BB : L.blocks()) 1428 IL->getBlocksSet().erase(BB); 1429 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) { 1430 return BB == PH || L.contains(BB); 1431 }); 1432 } 1433 1434 LI.changeLoopFor(PH, ParentL); 1435 L.getParentLoop()->removeChildLoop(&L); 1436 if (ParentL) 1437 ParentL->addChildLoop(&L); 1438 else 1439 LI.addTopLevelLoop(&L); 1440 } 1441 1442 // Now we update all the blocks which are no longer within the loop. 1443 auto &Blocks = L.getBlocksVector(); 1444 auto BlocksSplitI = 1445 LoopBlockSet.empty() 1446 ? Blocks.begin() 1447 : std::stable_partition( 1448 Blocks.begin(), Blocks.end(), 1449 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); }); 1450 1451 // Before we erase the list of unlooped blocks, build a set of them. 1452 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end()); 1453 if (LoopBlockSet.empty()) 1454 UnloopedBlocks.insert(PH); 1455 1456 // Now erase these blocks from the loop. 1457 for (auto *BB : make_range(BlocksSplitI, Blocks.end())) 1458 L.getBlocksSet().erase(BB); 1459 Blocks.erase(BlocksSplitI, Blocks.end()); 1460 1461 // Sort the exits in ascending loop depth, we'll work backwards across these 1462 // to process them inside out. 1463 std::stable_sort(ExitsInLoops.begin(), ExitsInLoops.end(), 1464 [&](BasicBlock *LHS, BasicBlock *RHS) { 1465 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS); 1466 }); 1467 1468 // We'll build up a set for each exit loop. 1469 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks; 1470 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop. 1471 1472 auto RemoveUnloopedBlocksFromLoop = 1473 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) { 1474 for (auto *BB : UnloopedBlocks) 1475 L.getBlocksSet().erase(BB); 1476 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) { 1477 return UnloopedBlocks.count(BB); 1478 }); 1479 }; 1480 1481 SmallVector<BasicBlock *, 16> Worklist; 1482 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) { 1483 assert(Worklist.empty() && "Didn't clear worklist!"); 1484 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!"); 1485 1486 // Grab the next exit block, in decreasing loop depth order. 1487 BasicBlock *ExitBB = ExitsInLoops.pop_back_val(); 1488 Loop &ExitL = *LI.getLoopFor(ExitBB); 1489 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!"); 1490 1491 // Erase all of the unlooped blocks from the loops between the previous 1492 // exit loop and this exit loop. This works because the ExitInLoops list is 1493 // sorted in increasing order of loop depth and thus we visit loops in 1494 // decreasing order of loop depth. 1495 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop()) 1496 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1497 1498 // Walk the CFG back until we hit the cloned PH adding everything reachable 1499 // and in the unlooped set to this exit block's loop. 1500 Worklist.push_back(ExitBB); 1501 do { 1502 BasicBlock *BB = Worklist.pop_back_val(); 1503 // We can stop recursing at the cloned preheader (if we get there). 1504 if (BB == PH) 1505 continue; 1506 1507 for (BasicBlock *PredBB : predecessors(BB)) { 1508 // If this pred has already been moved to our set or is part of some 1509 // (inner) loop, no update needed. 1510 if (!UnloopedBlocks.erase(PredBB)) { 1511 assert((NewExitLoopBlocks.count(PredBB) || 1512 ExitL.contains(LI.getLoopFor(PredBB))) && 1513 "Predecessor not in a nested loop (or already visited)!"); 1514 continue; 1515 } 1516 1517 // We just insert into the loop set here. We'll add these blocks to the 1518 // exit loop after we build up the set in a deterministic order rather 1519 // than the predecessor-influenced visit order. 1520 bool Inserted = NewExitLoopBlocks.insert(PredBB).second; 1521 (void)Inserted; 1522 assert(Inserted && "Should only visit an unlooped block once!"); 1523 1524 // And recurse through to its predecessors. 1525 Worklist.push_back(PredBB); 1526 } 1527 } while (!Worklist.empty()); 1528 1529 // If blocks in this exit loop were directly part of the original loop (as 1530 // opposed to a child loop) update the map to point to this exit loop. This 1531 // just updates a map and so the fact that the order is unstable is fine. 1532 for (auto *BB : NewExitLoopBlocks) 1533 if (Loop *BBL = LI.getLoopFor(BB)) 1534 if (BBL == &L || !L.contains(BBL)) 1535 LI.changeLoopFor(BB, &ExitL); 1536 1537 // We will remove the remaining unlooped blocks from this loop in the next 1538 // iteration or below. 1539 NewExitLoopBlocks.clear(); 1540 } 1541 1542 // Any remaining unlooped blocks are no longer part of any loop unless they 1543 // are part of some child loop. 1544 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop()) 1545 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); 1546 for (auto *BB : UnloopedBlocks) 1547 if (Loop *BBL = LI.getLoopFor(BB)) 1548 if (BBL == &L || !L.contains(BBL)) 1549 LI.changeLoopFor(BB, nullptr); 1550 1551 // Sink all the child loops whose headers are no longer in the loop set to 1552 // the parent (or to be top level loops). We reach into the loop and directly 1553 // update its subloop vector to make this batch update efficient. 1554 auto &SubLoops = L.getSubLoopsVector(); 1555 auto SubLoopsSplitI = 1556 LoopBlockSet.empty() 1557 ? SubLoops.begin() 1558 : std::stable_partition( 1559 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) { 1560 return LoopBlockSet.count(SubL->getHeader()); 1561 }); 1562 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) { 1563 HoistedLoops.push_back(HoistedL); 1564 HoistedL->setParentLoop(nullptr); 1565 1566 // To compute the new parent of this hoisted loop we look at where we 1567 // placed the preheader above. We can't lookup the header itself because we 1568 // retained the mapping from the header to the hoisted loop. But the 1569 // preheader and header should have the exact same new parent computed 1570 // based on the set of exit blocks from the original loop as the preheader 1571 // is a predecessor of the header and so reached in the reverse walk. And 1572 // because the loops were all in simplified form the preheader of the 1573 // hoisted loop can't be part of some *other* loop. 1574 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader())) 1575 NewParentL->addChildLoop(HoistedL); 1576 else 1577 LI.addTopLevelLoop(HoistedL); 1578 } 1579 SubLoops.erase(SubLoopsSplitI, SubLoops.end()); 1580 1581 // Actually delete the loop if nothing remained within it. 1582 if (Blocks.empty()) { 1583 assert(SubLoops.empty() && 1584 "Failed to remove all subloops from the original loop!"); 1585 if (Loop *ParentL = L.getParentLoop()) 1586 ParentL->removeChildLoop(llvm::find(*ParentL, &L)); 1587 else 1588 LI.removeLoop(llvm::find(LI, &L)); 1589 LI.destroy(&L); 1590 return false; 1591 } 1592 1593 return true; 1594 } 1595 1596 /// Helper to visit a dominator subtree, invoking a callable on each node. 1597 /// 1598 /// Returning false at any point will stop walking past that node of the tree. 1599 template <typename CallableT> 1600 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) { 1601 SmallVector<DomTreeNode *, 4> DomWorklist; 1602 DomWorklist.push_back(DT[BB]); 1603 #ifndef NDEBUG 1604 SmallPtrSet<DomTreeNode *, 4> Visited; 1605 Visited.insert(DT[BB]); 1606 #endif 1607 do { 1608 DomTreeNode *N = DomWorklist.pop_back_val(); 1609 1610 // Visit this node. 1611 if (!Callable(N->getBlock())) 1612 continue; 1613 1614 // Accumulate the child nodes. 1615 for (DomTreeNode *ChildN : *N) { 1616 assert(Visited.insert(ChildN).second && 1617 "Cannot visit a node twice when walking a tree!"); 1618 DomWorklist.push_back(ChildN); 1619 } 1620 } while (!DomWorklist.empty()); 1621 } 1622 1623 /// Take an invariant branch that has been determined to be safe and worthwhile 1624 /// to unswitch despite being non-trivial to do so and perform the unswitch. 1625 /// 1626 /// This directly updates the CFG to hoist the predicate out of the loop, and 1627 /// clone the necessary parts of the loop to maintain behavior. 1628 /// 1629 /// It also updates both dominator tree and loopinfo based on the unswitching. 1630 /// 1631 /// Once unswitching has been performed it runs the provided callback to report 1632 /// the new loops and no-longer valid loops to the caller. 1633 static bool unswitchInvariantBranch( 1634 Loop &L, BranchInst &BI, DominatorTree &DT, LoopInfo &LI, 1635 AssumptionCache &AC, 1636 function_ref<void(bool, ArrayRef<Loop *>)> NonTrivialUnswitchCB) { 1637 assert(BI.isConditional() && "Can only unswitch a conditional branch!"); 1638 assert(L.isLoopInvariant(BI.getCondition()) && 1639 "Can only unswitch an invariant branch condition!"); 1640 1641 // Constant and BBs tracking the cloned and continuing successor. 1642 const int ClonedSucc = 0; 1643 auto *ParentBB = BI.getParent(); 1644 auto *UnswitchedSuccBB = BI.getSuccessor(ClonedSucc); 1645 auto *ContinueSuccBB = BI.getSuccessor(1 - ClonedSucc); 1646 1647 assert(UnswitchedSuccBB != ContinueSuccBB && 1648 "Should not unswitch a branch that always goes to the same place!"); 1649 1650 // The branch should be in this exact loop. Any inner loop's invariant branch 1651 // should be handled by unswitching that inner loop. The caller of this 1652 // routine should filter out any candidates that remain (but were skipped for 1653 // whatever reason). 1654 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!"); 1655 1656 SmallVector<BasicBlock *, 4> ExitBlocks; 1657 L.getUniqueExitBlocks(ExitBlocks); 1658 1659 // We cannot unswitch if exit blocks contain a cleanuppad instruction as we 1660 // don't know how to split those exit blocks. 1661 // FIXME: We should teach SplitBlock to handle this and remove this 1662 // restriction. 1663 for (auto *ExitBB : ExitBlocks) 1664 if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) 1665 return false; 1666 1667 SmallPtrSet<BasicBlock *, 4> ExitBlockSet(ExitBlocks.begin(), 1668 ExitBlocks.end()); 1669 1670 // Compute the parent loop now before we start hacking on things. 1671 Loop *ParentL = L.getParentLoop(); 1672 1673 // Compute the outer-most loop containing one of our exit blocks. This is the 1674 // furthest up our loopnest which can be mutated, which we will use below to 1675 // update things. 1676 Loop *OuterExitL = &L; 1677 for (auto *ExitBB : ExitBlocks) { 1678 Loop *NewOuterExitL = LI.getLoopFor(ExitBB); 1679 if (!NewOuterExitL) { 1680 // We exited the entire nest with this block, so we're done. 1681 OuterExitL = nullptr; 1682 break; 1683 } 1684 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL)) 1685 OuterExitL = NewOuterExitL; 1686 } 1687 1688 // If the edge we *aren't* cloning in the unswitch (the continuing edge) 1689 // dominates its target, we can skip cloning the dominated region of the loop 1690 // and its exits. We compute this as a set of nodes to be skipped. 1691 SmallPtrSet<BasicBlock *, 4> SkippedLoopAndExitBlocks; 1692 if (ContinueSuccBB->getUniquePredecessor() || 1693 llvm::all_of(predecessors(ContinueSuccBB), [&](BasicBlock *PredBB) { 1694 return PredBB == ParentBB || DT.dominates(ContinueSuccBB, PredBB); 1695 })) { 1696 visitDomSubTree(DT, ContinueSuccBB, [&](BasicBlock *BB) { 1697 SkippedLoopAndExitBlocks.insert(BB); 1698 return true; 1699 }); 1700 } 1701 // Similarly, if the edge we *are* cloning in the unswitch (the unswitched 1702 // edge) dominates its target, we will end up with dead nodes in the original 1703 // loop and its exits that will need to be deleted. Here, we just retain that 1704 // the property holds and will compute the deleted set later. 1705 bool DeleteUnswitchedSucc = 1706 UnswitchedSuccBB->getUniquePredecessor() || 1707 llvm::all_of(predecessors(UnswitchedSuccBB), [&](BasicBlock *PredBB) { 1708 return PredBB == ParentBB || DT.dominates(UnswitchedSuccBB, PredBB); 1709 }); 1710 1711 // Split the preheader, so that we know that there is a safe place to insert 1712 // the conditional branch. We will change the preheader to have a conditional 1713 // branch on LoopCond. The original preheader will become the split point 1714 // between the unswitched versions, and we will have a new preheader for the 1715 // original loop. 1716 BasicBlock *SplitBB = L.getLoopPreheader(); 1717 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI); 1718 1719 // Keep a mapping for the cloned values. 1720 ValueToValueMapTy VMap; 1721 1722 // Keep track of the dominator tree updates needed. 1723 SmallVector<DominatorTree::UpdateType, 4> DTUpdates; 1724 1725 // Build the cloned blocks from the loop. 1726 auto *ClonedPH = buildClonedLoopBlocks( 1727 L, LoopPH, SplitBB, ExitBlocks, ParentBB, UnswitchedSuccBB, 1728 ContinueSuccBB, SkippedLoopAndExitBlocks, VMap, DTUpdates, AC, DT, LI); 1729 1730 // Remove the parent as a predecessor of the unswitched successor. 1731 UnswitchedSuccBB->removePredecessor(ParentBB, /*DontDeleteUselessPHIs*/ true); 1732 1733 // Now splice the branch from the original loop and use it to select between 1734 // the two loops. 1735 SplitBB->getTerminator()->eraseFromParent(); 1736 SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), BI); 1737 BI.setSuccessor(ClonedSucc, ClonedPH); 1738 BI.setSuccessor(1 - ClonedSucc, LoopPH); 1739 1740 // Create a new unconditional branch to the continuing block (as opposed to 1741 // the one cloned). 1742 BranchInst::Create(ContinueSuccBB, ParentBB); 1743 1744 // Before we update the dominator tree, collect the dead blocks if we're going 1745 // to end up deleting the unswitched successor. 1746 SmallVector<BasicBlock *, 16> DeadBlocks; 1747 if (DeleteUnswitchedSucc) { 1748 DeadBlocks.push_back(UnswitchedSuccBB); 1749 for (int i = 0; i < (int)DeadBlocks.size(); ++i) { 1750 // If we reach an exit block, stop recursing as the unswitched loop will 1751 // end up reaching the merge block which we make the successor of the 1752 // exit. 1753 if (ExitBlockSet.count(DeadBlocks[i])) 1754 continue; 1755 1756 // Insert the children that are within the loop or exit block set. Other 1757 // children may reach out of the loop. While we don't expect these to be 1758 // dead (as the unswitched clone should reach them) we don't try to prove 1759 // that here. 1760 for (DomTreeNode *ChildN : *DT[DeadBlocks[i]]) 1761 if (L.contains(ChildN->getBlock()) || 1762 ExitBlockSet.count(ChildN->getBlock())) 1763 DeadBlocks.push_back(ChildN->getBlock()); 1764 } 1765 } 1766 1767 // Add the remaining edges to our updates and apply them to get an up-to-date 1768 // dominator tree. Note that this will cause the dead blocks above to be 1769 // unreachable and no longer in the dominator tree. 1770 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB}); 1771 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); 1772 DT.applyUpdates(DTUpdates); 1773 1774 // Build the cloned loop structure itself. This may be substantially 1775 // different from the original structure due to the simplified CFG. This also 1776 // handles inserting all the cloned blocks into the correct loops. 1777 SmallVector<Loop *, 4> NonChildClonedLoops; 1778 Loop *ClonedL = 1779 buildClonedLoops(L, ExitBlocks, VMap, LI, NonChildClonedLoops); 1780 1781 // Delete anything that was made dead in the original loop due to 1782 // unswitching. 1783 if (!DeadBlocks.empty()) 1784 deleteDeadBlocksFromLoop(L, DeadBlocks, ExitBlocks, DT, LI); 1785 1786 SmallVector<Loop *, 4> HoistedLoops; 1787 bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops); 1788 1789 // This transformation has a high risk of corrupting the dominator tree, and 1790 // the below steps to rebuild loop structures will result in hard to debug 1791 // errors in that case so verify that the dominator tree is sane first. 1792 // FIXME: Remove this when the bugs stop showing up and rely on existing 1793 // verification steps. 1794 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 1795 1796 // We can change which blocks are exit blocks of all the cloned sibling 1797 // loops, the current loop, and any parent loops which shared exit blocks 1798 // with the current loop. As a consequence, we need to re-form LCSSA for 1799 // them. But we shouldn't need to re-form LCSSA for any child loops. 1800 // FIXME: This could be made more efficient by tracking which exit blocks are 1801 // new, and focusing on them, but that isn't likely to be necessary. 1802 // 1803 // In order to reasonably rebuild LCSSA we need to walk inside-out across the 1804 // loop nest and update every loop that could have had its exits changed. We 1805 // also need to cover any intervening loops. We add all of these loops to 1806 // a list and sort them by loop depth to achieve this without updating 1807 // unnecessary loops. 1808 auto UpdateLCSSA = [&](Loop &UpdateL) { 1809 #ifndef NDEBUG 1810 UpdateL.verifyLoop(); 1811 for (Loop *ChildL : UpdateL) { 1812 ChildL->verifyLoop(); 1813 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) && 1814 "Perturbed a child loop's LCSSA form!"); 1815 } 1816 #endif 1817 formLCSSA(UpdateL, DT, &LI, nullptr); 1818 }; 1819 1820 // For non-child cloned loops and hoisted loops, we just need to update LCSSA 1821 // and we can do it in any order as they don't nest relative to each other. 1822 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) 1823 UpdateLCSSA(*UpdatedL); 1824 1825 // If the original loop had exit blocks, walk up through the outer most loop 1826 // of those exit blocks to update LCSSA and form updated dedicated exits. 1827 if (OuterExitL != &L) { 1828 SmallVector<Loop *, 4> OuterLoops; 1829 // We start with the cloned loop and the current loop if they are loops and 1830 // move toward OuterExitL. Also, if either the cloned loop or the current 1831 // loop have become top level loops we need to walk all the way out. 1832 if (ClonedL) { 1833 OuterLoops.push_back(ClonedL); 1834 if (!ClonedL->getParentLoop()) 1835 OuterExitL = nullptr; 1836 } 1837 if (IsStillLoop) { 1838 OuterLoops.push_back(&L); 1839 if (!L.getParentLoop()) 1840 OuterExitL = nullptr; 1841 } 1842 // Grab all of the enclosing loops now. 1843 for (Loop *OuterL = ParentL; OuterL != OuterExitL; 1844 OuterL = OuterL->getParentLoop()) 1845 OuterLoops.push_back(OuterL); 1846 1847 // Finally, update our list of outer loops. This is nicely ordered to work 1848 // inside-out. 1849 for (Loop *OuterL : OuterLoops) { 1850 // First build LCSSA for this loop so that we can preserve it when 1851 // forming dedicated exits. We don't want to perturb some other loop's 1852 // LCSSA while doing that CFG edit. 1853 UpdateLCSSA(*OuterL); 1854 1855 // For loops reached by this loop's original exit blocks we may 1856 // introduced new, non-dedicated exits. At least try to re-form dedicated 1857 // exits for these loops. This may fail if they couldn't have dedicated 1858 // exits to start with. 1859 formDedicatedExitBlocks(OuterL, &DT, &LI, /*PreserveLCSSA*/ true); 1860 } 1861 } 1862 1863 #ifndef NDEBUG 1864 // Verify the entire loop structure to catch any incorrect updates before we 1865 // progress in the pass pipeline. 1866 LI.verify(DT); 1867 #endif 1868 1869 // Now that we've unswitched something, make callbacks to report the changes. 1870 // For that we need to merge together the updated loops and the cloned loops 1871 // and check whether the original loop survived. 1872 SmallVector<Loop *, 4> SibLoops; 1873 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) 1874 if (UpdatedL->getParentLoop() == ParentL) 1875 SibLoops.push_back(UpdatedL); 1876 NonTrivialUnswitchCB(IsStillLoop, SibLoops); 1877 1878 ++NumBranches; 1879 return true; 1880 } 1881 1882 /// Recursively compute the cost of a dominator subtree based on the per-block 1883 /// cost map provided. 1884 /// 1885 /// The recursive computation is memozied into the provided DT-indexed cost map 1886 /// to allow querying it for most nodes in the domtree without it becoming 1887 /// quadratic. 1888 static int 1889 computeDomSubtreeCost(DomTreeNode &N, 1890 const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap, 1891 SmallDenseMap<DomTreeNode *, int, 4> &DTCostMap) { 1892 // Don't accumulate cost (or recurse through) blocks not in our block cost 1893 // map and thus not part of the duplication cost being considered. 1894 auto BBCostIt = BBCostMap.find(N.getBlock()); 1895 if (BBCostIt == BBCostMap.end()) 1896 return 0; 1897 1898 // Lookup this node to see if we already computed its cost. 1899 auto DTCostIt = DTCostMap.find(&N); 1900 if (DTCostIt != DTCostMap.end()) 1901 return DTCostIt->second; 1902 1903 // If not, we have to compute it. We can't use insert above and update 1904 // because computing the cost may insert more things into the map. 1905 int Cost = std::accumulate( 1906 N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) { 1907 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap); 1908 }); 1909 bool Inserted = DTCostMap.insert({&N, Cost}).second; 1910 (void)Inserted; 1911 assert(Inserted && "Should not insert a node while visiting children!"); 1912 return Cost; 1913 } 1914 1915 /// Unswitch control flow predicated on loop invariant conditions. 1916 /// 1917 /// This first hoists all branches or switches which are trivial (IE, do not 1918 /// require duplicating any part of the loop) out of the loop body. It then 1919 /// looks at other loop invariant control flows and tries to unswitch those as 1920 /// well by cloning the loop if the result is small enough. 1921 static bool 1922 unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, 1923 TargetTransformInfo &TTI, bool NonTrivial, 1924 function_ref<void(bool, ArrayRef<Loop *>)> NonTrivialUnswitchCB) { 1925 assert(L.isRecursivelyLCSSAForm(DT, LI) && 1926 "Loops must be in LCSSA form before unswitching."); 1927 bool Changed = false; 1928 1929 // Must be in loop simplified form: we need a preheader and dedicated exits. 1930 if (!L.isLoopSimplifyForm()) 1931 return false; 1932 1933 // Try trivial unswitch first before loop over other basic blocks in the loop. 1934 Changed |= unswitchAllTrivialConditions(L, DT, LI); 1935 1936 // If we're not doing non-trivial unswitching, we're done. We both accept 1937 // a parameter but also check a local flag that can be used for testing 1938 // a debugging. 1939 if (!NonTrivial && !EnableNonTrivialUnswitch) 1940 return Changed; 1941 1942 // Collect all remaining invariant branch conditions within this loop (as 1943 // opposed to an inner loop which would be handled when visiting that inner 1944 // loop). 1945 SmallVector<TerminatorInst *, 4> UnswitchCandidates; 1946 for (auto *BB : L.blocks()) 1947 if (LI.getLoopFor(BB) == &L) 1948 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1949 if (BI->isConditional() && L.isLoopInvariant(BI->getCondition()) && 1950 BI->getSuccessor(0) != BI->getSuccessor(1)) 1951 UnswitchCandidates.push_back(BI); 1952 1953 // If we didn't find any candidates, we're done. 1954 if (UnswitchCandidates.empty()) 1955 return Changed; 1956 1957 // Check if there are irreducible CFG cycles in this loop. If so, we cannot 1958 // easily unswitch non-trivial edges out of the loop. Doing so might turn the 1959 // irreducible control flow into reducible control flow and introduce new 1960 // loops "out of thin air". If we ever discover important use cases for doing 1961 // this, we can add support to loop unswitch, but it is a lot of complexity 1962 // for what seems little or no real world benifit. 1963 LoopBlocksRPO RPOT(&L); 1964 RPOT.perform(&LI); 1965 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 1966 return Changed; 1967 1968 DEBUG(dbgs() << "Considering " << UnswitchCandidates.size() 1969 << " non-trivial loop invariant conditions for unswitching.\n"); 1970 1971 // Given that unswitching these terminators will require duplicating parts of 1972 // the loop, so we need to be able to model that cost. Compute the ephemeral 1973 // values and set up a data structure to hold per-BB costs. We cache each 1974 // block's cost so that we don't recompute this when considering different 1975 // subsets of the loop for duplication during unswitching. 1976 SmallPtrSet<const Value *, 4> EphValues; 1977 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues); 1978 SmallDenseMap<BasicBlock *, int, 4> BBCostMap; 1979 1980 // Compute the cost of each block, as well as the total loop cost. Also, bail 1981 // out if we see instructions which are incompatible with loop unswitching 1982 // (convergent, noduplicate, or cross-basic-block tokens). 1983 // FIXME: We might be able to safely handle some of these in non-duplicated 1984 // regions. 1985 int LoopCost = 0; 1986 for (auto *BB : L.blocks()) { 1987 int Cost = 0; 1988 for (auto &I : *BB) { 1989 if (EphValues.count(&I)) 1990 continue; 1991 1992 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB)) 1993 return Changed; 1994 if (auto CS = CallSite(&I)) 1995 if (CS.isConvergent() || CS.cannotDuplicate()) 1996 return Changed; 1997 1998 Cost += TTI.getUserCost(&I); 1999 } 2000 assert(Cost >= 0 && "Must not have negative costs!"); 2001 LoopCost += Cost; 2002 assert(LoopCost >= 0 && "Must not have negative loop costs!"); 2003 BBCostMap[BB] = Cost; 2004 } 2005 DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n"); 2006 2007 // Now we find the best candidate by searching for the one with the following 2008 // properties in order: 2009 // 2010 // 1) An unswitching cost below the threshold 2011 // 2) The smallest number of duplicated unswitch candidates (to avoid 2012 // creating redundant subsequent unswitching) 2013 // 3) The smallest cost after unswitching. 2014 // 2015 // We prioritize reducing fanout of unswitch candidates provided the cost 2016 // remains below the threshold because this has a multiplicative effect. 2017 // 2018 // This requires memoizing each dominator subtree to avoid redundant work. 2019 // 2020 // FIXME: Need to actually do the number of candidates part above. 2021 SmallDenseMap<DomTreeNode *, int, 4> DTCostMap; 2022 // Given a terminator which might be unswitched, computes the non-duplicated 2023 // cost for that terminator. 2024 auto ComputeUnswitchedCost = [&](TerminatorInst *TI) { 2025 BasicBlock &BB = *TI->getParent(); 2026 SmallPtrSet<BasicBlock *, 4> Visited; 2027 2028 int Cost = LoopCost; 2029 for (BasicBlock *SuccBB : successors(&BB)) { 2030 // Don't count successors more than once. 2031 if (!Visited.insert(SuccBB).second) 2032 continue; 2033 2034 // This successor's domtree will not need to be duplicated after 2035 // unswitching if the edge to the successor dominates it (and thus the 2036 // entire tree). This essentially means there is no other path into this 2037 // subtree and so it will end up live in only one clone of the loop. 2038 if (SuccBB->getUniquePredecessor() || 2039 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { 2040 return PredBB == &BB || DT.dominates(SuccBB, PredBB); 2041 })) { 2042 Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap); 2043 assert(Cost >= 0 && 2044 "Non-duplicated cost should never exceed total loop cost!"); 2045 } 2046 } 2047 2048 // Now scale the cost by the number of unique successors minus one. We 2049 // subtract one because there is already at least one copy of the entire 2050 // loop. This is computing the new cost of unswitching a condition. 2051 assert(Visited.size() > 1 && 2052 "Cannot unswitch a condition without multiple distinct successors!"); 2053 return Cost * (Visited.size() - 1); 2054 }; 2055 TerminatorInst *BestUnswitchTI = nullptr; 2056 int BestUnswitchCost; 2057 for (TerminatorInst *CandidateTI : UnswitchCandidates) { 2058 int CandidateCost = ComputeUnswitchedCost(CandidateTI); 2059 DEBUG(dbgs() << " Computed cost of " << CandidateCost 2060 << " for unswitch candidate: " << *CandidateTI << "\n"); 2061 if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) { 2062 BestUnswitchTI = CandidateTI; 2063 BestUnswitchCost = CandidateCost; 2064 } 2065 } 2066 2067 if (BestUnswitchCost < UnswitchThreshold) { 2068 DEBUG(dbgs() << " Trying to unswitch non-trivial (cost = " 2069 << BestUnswitchCost << ") branch: " << *BestUnswitchTI 2070 << "\n"); 2071 Changed |= unswitchInvariantBranch(L, cast<BranchInst>(*BestUnswitchTI), DT, 2072 LI, AC, NonTrivialUnswitchCB); 2073 } else { 2074 DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << BestUnswitchCost 2075 << "\n"); 2076 } 2077 2078 return Changed; 2079 } 2080 2081 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, 2082 LoopStandardAnalysisResults &AR, 2083 LPMUpdater &U) { 2084 Function &F = *L.getHeader()->getParent(); 2085 (void)F; 2086 2087 DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L << "\n"); 2088 2089 // Save the current loop name in a variable so that we can report it even 2090 // after it has been deleted. 2091 std::string LoopName = L.getName(); 2092 2093 auto NonTrivialUnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid, 2094 ArrayRef<Loop *> NewLoops) { 2095 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2096 U.addSiblingLoops(NewLoops); 2097 2098 // If the current loop remains valid, we should revisit it to catch any 2099 // other unswitch opportunities. Otherwise, we need to mark it as deleted. 2100 if (CurrentLoopValid) 2101 U.revisitCurrentLoop(); 2102 else 2103 U.markLoopAsDeleted(L, LoopName); 2104 }; 2105 2106 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, 2107 NonTrivialUnswitchCB)) 2108 return PreservedAnalyses::all(); 2109 2110 // Historically this pass has had issues with the dominator tree so verify it 2111 // in asserts builds. 2112 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast)); 2113 return getLoopPassPreservedAnalyses(); 2114 } 2115 2116 namespace { 2117 2118 class SimpleLoopUnswitchLegacyPass : public LoopPass { 2119 bool NonTrivial; 2120 2121 public: 2122 static char ID; // Pass ID, replacement for typeid 2123 2124 explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false) 2125 : LoopPass(ID), NonTrivial(NonTrivial) { 2126 initializeSimpleLoopUnswitchLegacyPassPass( 2127 *PassRegistry::getPassRegistry()); 2128 } 2129 2130 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 2131 2132 void getAnalysisUsage(AnalysisUsage &AU) const override { 2133 AU.addRequired<AssumptionCacheTracker>(); 2134 AU.addRequired<TargetTransformInfoWrapperPass>(); 2135 getLoopAnalysisUsage(AU); 2136 } 2137 }; 2138 2139 } // end anonymous namespace 2140 2141 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { 2142 if (skipLoop(L)) 2143 return false; 2144 2145 Function &F = *L->getHeader()->getParent(); 2146 2147 DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L << "\n"); 2148 2149 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2150 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2151 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 2152 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 2153 2154 auto NonTrivialUnswitchCB = [&L, &LPM](bool CurrentLoopValid, 2155 ArrayRef<Loop *> NewLoops) { 2156 // If we did a non-trivial unswitch, we have added new (cloned) loops. 2157 for (auto *NewL : NewLoops) 2158 LPM.addLoop(*NewL); 2159 2160 // If the current loop remains valid, re-add it to the queue. This is 2161 // a little wasteful as we'll finish processing the current loop as well, 2162 // but it is the best we can do in the old PM. 2163 if (CurrentLoopValid) 2164 LPM.addLoop(*L); 2165 else 2166 LPM.markLoopAsDeleted(*L); 2167 }; 2168 2169 bool Changed = 2170 unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, NonTrivialUnswitchCB); 2171 2172 // If anything was unswitched, also clear any cached information about this 2173 // loop. 2174 LPM.deleteSimpleAnalysisLoop(L); 2175 2176 // Historically this pass has had issues with the dominator tree so verify it 2177 // in asserts builds. 2178 assert(DT.verify(DominatorTree::VerificationLevel::Fast)); 2179 2180 return Changed; 2181 } 2182 2183 char SimpleLoopUnswitchLegacyPass::ID = 0; 2184 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", 2185 "Simple unswitch loops", false, false) 2186 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 2187 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 2188 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 2189 INITIALIZE_PASS_DEPENDENCY(LoopPass) 2190 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 2191 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", 2192 "Simple unswitch loops", false, false) 2193 2194 Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) { 2195 return new SimpleLoopUnswitchLegacyPass(NonTrivial); 2196 } 2197