1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===// 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 // This pass transforms loops that contain branches on loop-invariant conditions 11 // to have multiple loops. For example, it turns the left into the right code: 12 // 13 // for (...) if (lic) 14 // A for (...) 15 // if (lic) A; B; C 16 // B else 17 // C for (...) 18 // A; C 19 // 20 // This can increase the size of the code exponentially (doubling it every time 21 // a loop is unswitched) so we only unswitch if the resultant code will be 22 // smaller than a threshold. 23 // 24 // This pass expects LICM to be run before it to hoist invariant conditions out 25 // of the loop, to make the unswitching opportunity obvious. 26 // 27 //===----------------------------------------------------------------------===// 28 29 #define DEBUG_TYPE "loop-unswitch" 30 #include "llvm/Transforms/Scalar.h" 31 #include "llvm/Constants.h" 32 #include "llvm/DerivedTypes.h" 33 #include "llvm/Function.h" 34 #include "llvm/Instructions.h" 35 #include "llvm/Analysis/ConstantFolding.h" 36 #include "llvm/Analysis/InlineCost.h" 37 #include "llvm/Analysis/InstructionSimplify.h" 38 #include "llvm/Analysis/LoopInfo.h" 39 #include "llvm/Analysis/LoopPass.h" 40 #include "llvm/Analysis/Dominators.h" 41 #include "llvm/Transforms/Utils/Cloning.h" 42 #include "llvm/Transforms/Utils/Local.h" 43 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 44 #include "llvm/ADT/Statistic.h" 45 #include "llvm/ADT/SmallPtrSet.h" 46 #include "llvm/ADT/STLExtras.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/Debug.h" 49 #include "llvm/Support/raw_ostream.h" 50 #include <algorithm> 51 #include <set> 52 using namespace llvm; 53 54 STATISTIC(NumBranches, "Number of branches unswitched"); 55 STATISTIC(NumSwitches, "Number of switches unswitched"); 56 STATISTIC(NumSelects , "Number of selects unswitched"); 57 STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 58 STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 59 60 // The specific value of 50 here was chosen based only on intuition and a 61 // few specific examples. 62 static cl::opt<unsigned> 63 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 64 cl::init(50), cl::Hidden); 65 66 namespace { 67 class LoopUnswitch : public LoopPass { 68 LoopInfo *LI; // Loop information 69 LPPassManager *LPM; 70 71 // LoopProcessWorklist - Used to check if second loop needs processing 72 // after RewriteLoopBodyWithConditionConstant rewrites first loop. 73 std::vector<Loop*> LoopProcessWorklist; 74 SmallPtrSet<Value *,8> UnswitchedVals; 75 76 bool OptimizeForSize; 77 bool redoLoop; 78 79 Loop *currentLoop; 80 DominanceFrontier *DF; 81 DominatorTree *DT; 82 BasicBlock *loopHeader; 83 BasicBlock *loopPreheader; 84 85 // LoopBlocks contains all of the basic blocks of the loop, including the 86 // preheader of the loop, the body of the loop, and the exit blocks of the 87 // loop, in that order. 88 std::vector<BasicBlock*> LoopBlocks; 89 // NewBlocks contained cloned copy of basic blocks from LoopBlocks. 90 std::vector<BasicBlock*> NewBlocks; 91 92 public: 93 static char ID; // Pass ID, replacement for typeid 94 explicit LoopUnswitch(bool Os = false) : 95 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false), 96 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL), 97 loopPreheader(NULL) {} 98 99 bool runOnLoop(Loop *L, LPPassManager &LPM); 100 bool processCurrentLoop(); 101 102 /// This transformation requires natural loop information & requires that 103 /// loop preheaders be inserted into the CFG... 104 /// 105 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 106 AU.addRequiredID(LoopSimplifyID); 107 AU.addPreservedID(LoopSimplifyID); 108 AU.addRequired<LoopInfo>(); 109 AU.addPreserved<LoopInfo>(); 110 AU.addRequiredID(LCSSAID); 111 AU.addPreservedID(LCSSAID); 112 AU.addPreserved<DominatorTree>(); 113 AU.addPreserved<DominanceFrontier>(); 114 } 115 116 private: 117 118 virtual void releaseMemory() { 119 UnswitchedVals.clear(); 120 } 121 122 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 123 /// remove it. 124 void RemoveLoopFromWorklist(Loop *L) { 125 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 126 LoopProcessWorklist.end(), L); 127 if (I != LoopProcessWorklist.end()) 128 LoopProcessWorklist.erase(I); 129 } 130 131 void initLoopData() { 132 loopHeader = currentLoop->getHeader(); 133 loopPreheader = currentLoop->getLoopPreheader(); 134 } 135 136 /// Split all of the edges from inside the loop to their exit blocks. 137 /// Update the appropriate Phi nodes as we do so. 138 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks); 139 140 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); 141 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 142 BasicBlock *ExitBlock); 143 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 144 145 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 146 Constant *Val, bool isEqual); 147 148 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 149 BasicBlock *TrueDest, 150 BasicBlock *FalseDest, 151 Instruction *InsertPt); 152 153 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 154 void RemoveBlockIfDead(BasicBlock *BB, 155 std::vector<Instruction*> &Worklist, Loop *l); 156 void RemoveLoopFromHierarchy(Loop *L); 157 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, 158 BasicBlock **LoopExit = 0); 159 160 }; 161 } 162 char LoopUnswitch::ID = 0; 163 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 164 165 Pass *llvm::createLoopUnswitchPass(bool Os) { 166 return new LoopUnswitch(Os); 167 } 168 169 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 170 /// invariant in the loop, or has an invariant piece, return the invariant. 171 /// Otherwise, return null. 172 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 173 // We can never unswitch on vector conditions. 174 if (Cond->getType()->isVectorTy()) 175 return 0; 176 177 // Constants should be folded, not unswitched on! 178 if (isa<Constant>(Cond)) return 0; 179 180 // TODO: Handle: br (VARIANT|INVARIANT). 181 182 // Hoist simple values out. 183 if (L->makeLoopInvariant(Cond, Changed)) 184 return Cond; 185 186 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 187 if (BO->getOpcode() == Instruction::And || 188 BO->getOpcode() == Instruction::Or) { 189 // If either the left or right side is invariant, we can unswitch on this, 190 // which will cause the branch to go away in one loop and the condition to 191 // simplify in the other one. 192 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 193 return LHS; 194 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 195 return RHS; 196 } 197 198 return 0; 199 } 200 201 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 202 LI = &getAnalysis<LoopInfo>(); 203 LPM = &LPM_Ref; 204 DF = getAnalysisIfAvailable<DominanceFrontier>(); 205 DT = getAnalysisIfAvailable<DominatorTree>(); 206 currentLoop = L; 207 Function *F = currentLoop->getHeader()->getParent(); 208 bool Changed = false; 209 do { 210 assert(currentLoop->isLCSSAForm(*DT)); 211 redoLoop = false; 212 Changed |= processCurrentLoop(); 213 } while(redoLoop); 214 215 if (Changed) { 216 // FIXME: Reconstruct dom info, because it is not preserved properly. 217 if (DT) 218 DT->runOnFunction(*F); 219 if (DF) 220 DF->runOnFunction(*F); 221 } 222 return Changed; 223 } 224 225 /// processCurrentLoop - Do actual work and unswitch loop if possible 226 /// and profitable. 227 bool LoopUnswitch::processCurrentLoop() { 228 bool Changed = false; 229 LLVMContext &Context = currentLoop->getHeader()->getContext(); 230 231 // Loop over all of the basic blocks in the loop. If we find an interior 232 // block that is branching on a loop-invariant condition, we can unswitch this 233 // loop. 234 for (Loop::block_iterator I = currentLoop->block_begin(), 235 E = currentLoop->block_end(); I != E; ++I) { 236 TerminatorInst *TI = (*I)->getTerminator(); 237 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 238 // If this isn't branching on an invariant condition, we can't unswitch 239 // it. 240 if (BI->isConditional()) { 241 // See if this, or some part of it, is loop invariant. If so, we can 242 // unswitch on it if we desire. 243 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 244 currentLoop, Changed); 245 if (LoopCond && UnswitchIfProfitable(LoopCond, 246 ConstantInt::getTrue(Context))) { 247 ++NumBranches; 248 return true; 249 } 250 } 251 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 252 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 253 currentLoop, Changed); 254 if (LoopCond && SI->getNumCases() > 1) { 255 // Find a value to unswitch on: 256 // FIXME: this should chose the most expensive case! 257 Constant *UnswitchVal = SI->getCaseValue(1); 258 // Do not process same value again and again. 259 if (!UnswitchedVals.insert(UnswitchVal)) 260 continue; 261 262 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 263 ++NumSwitches; 264 return true; 265 } 266 } 267 } 268 269 // Scan the instructions to check for unswitchable values. 270 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 271 BBI != E; ++BBI) 272 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 273 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 274 currentLoop, Changed); 275 if (LoopCond && UnswitchIfProfitable(LoopCond, 276 ConstantInt::getTrue(Context))) { 277 ++NumSelects; 278 return true; 279 } 280 } 281 } 282 return Changed; 283 } 284 285 /// isTrivialLoopExitBlock - Check to see if all paths from BB either: 286 /// 1. Exit the loop with no side effects. 287 /// 2. Branch to the latch block with no side-effects. 288 /// 289 /// If these conditions are true, we return true and set ExitBB to the block we 290 /// exit through. 291 /// 292 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 293 BasicBlock *&ExitBB, 294 std::set<BasicBlock*> &Visited) { 295 if (!Visited.insert(BB).second) { 296 // Already visited and Ok, end of recursion. 297 return true; 298 } else if (!L->contains(BB)) { 299 // Otherwise, this is a loop exit, this is fine so long as this is the 300 // first exit. 301 if (ExitBB != 0) return false; 302 ExitBB = BB; 303 return true; 304 } 305 306 // Otherwise, this is an unvisited intra-loop node. Check all successors. 307 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 308 // Check to see if the successor is a trivial loop exit. 309 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 310 return false; 311 } 312 313 // Okay, everything after this looks good, check to make sure that this block 314 // doesn't include any side effects. 315 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 316 if (I->mayHaveSideEffects()) 317 return false; 318 319 return true; 320 } 321 322 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally 323 /// leads to an exit from the specified loop, and has no side-effects in the 324 /// process. If so, return the block that is exited to, otherwise return null. 325 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 326 std::set<BasicBlock*> Visited; 327 Visited.insert(L->getHeader()); // Branches to header are ok. 328 BasicBlock *ExitBB = 0; 329 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 330 return ExitBB; 331 return 0; 332 } 333 334 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 335 /// trivial: that is, that the condition controls whether or not the loop does 336 /// anything at all. If this is a trivial condition, unswitching produces no 337 /// code duplications (equivalently, it produces a simpler loop and a new empty 338 /// loop, which gets deleted). 339 /// 340 /// If this is a trivial condition, return true, otherwise return false. When 341 /// returning true, this sets Cond and Val to the condition that controls the 342 /// trivial condition: when Cond dynamically equals Val, the loop is known to 343 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when 344 /// Cond == Val. 345 /// 346 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 347 BasicBlock **LoopExit) { 348 BasicBlock *Header = currentLoop->getHeader(); 349 TerminatorInst *HeaderTerm = Header->getTerminator(); 350 LLVMContext &Context = Header->getContext(); 351 352 BasicBlock *LoopExitBB = 0; 353 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 354 // If the header block doesn't end with a conditional branch on Cond, we 355 // can't handle it. 356 if (!BI->isConditional() || BI->getCondition() != Cond) 357 return false; 358 359 // Check to see if a successor of the branch is guaranteed to go to the 360 // latch block or exit through a one exit block without having any 361 // side-effects. If so, determine the value of Cond that causes it to do 362 // this. 363 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 364 BI->getSuccessor(0)))) { 365 if (Val) *Val = ConstantInt::getTrue(Context); 366 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 367 BI->getSuccessor(1)))) { 368 if (Val) *Val = ConstantInt::getFalse(Context); 369 } 370 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 371 // If this isn't a switch on Cond, we can't handle it. 372 if (SI->getCondition() != Cond) return false; 373 374 // Check to see if a successor of the switch is guaranteed to go to the 375 // latch block or exit through a one exit block without having any 376 // side-effects. If so, determine the value of Cond that causes it to do 377 // this. Note that we can't trivially unswitch on the default case. 378 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 379 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 380 SI->getSuccessor(i)))) { 381 // Okay, we found a trivial case, remember the value that is trivial. 382 if (Val) *Val = SI->getCaseValue(i); 383 break; 384 } 385 } 386 387 // If we didn't find a single unique LoopExit block, or if the loop exit block 388 // contains phi nodes, this isn't trivial. 389 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 390 return false; // Can't handle this. 391 392 if (LoopExit) *LoopExit = LoopExitBB; 393 394 // We already know that nothing uses any scalar values defined inside of this 395 // loop. As such, we just have to check to see if this loop will execute any 396 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 397 // part of the loop that the code *would* execute. We already checked the 398 // tail, check the header now. 399 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 400 if (I->mayHaveSideEffects()) 401 return false; 402 return true; 403 } 404 405 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 406 /// LoopCond == Val to simplify the loop. If we decide that this is profitable, 407 /// unswitch the loop, reprocess the pieces, then return true. 408 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { 409 410 initLoopData(); 411 412 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 413 if (!loopPreheader) 414 return false; 415 416 Function *F = loopHeader->getParent(); 417 418 Constant *CondVal = 0; 419 BasicBlock *ExitBlock = 0; 420 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 421 // If the condition is trivial, always unswitch. There is no code growth 422 // for this case. 423 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 424 return true; 425 } 426 427 // Check to see if it would be profitable to unswitch current loop. 428 429 // Do not do non-trivial unswitch while optimizing for size. 430 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) 431 return false; 432 433 // FIXME: This is overly conservative because it does not take into 434 // consideration code simplification opportunities and code that can 435 // be shared by the resultant unswitched loops. 436 CodeMetrics Metrics; 437 for (Loop::block_iterator I = currentLoop->block_begin(), 438 E = currentLoop->block_end(); 439 I != E; ++I) 440 Metrics.analyzeBasicBlock(*I); 441 442 // Limit the number of instructions to avoid causing significant code 443 // expansion, and the number of basic blocks, to avoid loops with 444 // large numbers of branches which cause loop unswitching to go crazy. 445 // This is a very ad-hoc heuristic. 446 if (Metrics.NumInsts > Threshold || 447 Metrics.NumBlocks * 5 > Threshold || 448 Metrics.containsIndirectBr || Metrics.isRecursive) { 449 DEBUG(dbgs() << "NOT unswitching loop %" 450 << currentLoop->getHeader()->getName() << ", cost too high: " 451 << currentLoop->getBlocks().size() << "\n"); 452 return false; 453 } 454 455 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 456 return true; 457 } 458 459 // RemapInstruction - Convert the instruction operands from referencing the 460 // current values into those specified by VMap. 461 // 462 static inline void RemapInstruction(Instruction *I, 463 ValueMap<const Value *, Value*> &VMap) { 464 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 465 Value *Op = I->getOperand(op); 466 ValueMap<const Value *, Value*>::iterator It = VMap.find(Op); 467 if (It != VMap.end()) Op = It->second; 468 I->setOperand(op, Op); 469 } 470 } 471 472 /// CloneLoop - Recursively clone the specified loop and all of its children, 473 /// mapping the blocks with the specified map. 474 static Loop *CloneLoop(Loop *L, Loop *PL, ValueMap<const Value*, Value*> &VM, 475 LoopInfo *LI, LPPassManager *LPM) { 476 Loop *New = new Loop(); 477 LPM->insertLoop(New, PL); 478 479 // Add all of the blocks in L to the new loop. 480 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 481 I != E; ++I) 482 if (LI->getLoopFor(*I) == L) 483 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 484 485 // Add all of the subloops to the new loop. 486 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 487 CloneLoop(*I, New, VM, LI, LPM); 488 489 return New; 490 } 491 492 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 493 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 494 /// code immediately before InsertPt. 495 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 496 BasicBlock *TrueDest, 497 BasicBlock *FalseDest, 498 Instruction *InsertPt) { 499 // Insert a conditional branch on LIC to the two preheaders. The original 500 // code is the true version and the new code is the false version. 501 Value *BranchVal = LIC; 502 if (!isa<ConstantInt>(Val) || 503 Val->getType() != Type::getInt1Ty(LIC->getContext())) 504 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp"); 505 else if (Val != ConstantInt::getTrue(Val->getContext())) 506 // We want to enter the new loop when the condition is true. 507 std::swap(TrueDest, FalseDest); 508 509 // Insert the new branch. 510 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 511 512 // If either edge is critical, split it. This helps preserve LoopSimplify 513 // form for enclosing loops. 514 SplitCriticalEdge(BI, 0, this); 515 SplitCriticalEdge(BI, 1, this); 516 } 517 518 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 519 /// condition in it (a cond branch from its header block to its latch block, 520 /// where the path through the loop that doesn't execute its body has no 521 /// side-effects), unswitch it. This doesn't involve any code duplication, just 522 /// moving the conditional branch outside of the loop and updating loop info. 523 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 524 Constant *Val, 525 BasicBlock *ExitBlock) { 526 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 527 << loopHeader->getName() << " [" << L->getBlocks().size() 528 << " blocks] in Function " << L->getHeader()->getParent()->getName() 529 << " on cond: " << *Val << " == " << *Cond << "\n"); 530 531 // First step, split the preheader, so that we know that there is a safe place 532 // to insert the conditional branch. We will change loopPreheader to have a 533 // conditional branch on Cond. 534 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 535 536 // Now that we have a place to insert the conditional branch, create a place 537 // to branch to: this is the exit block out of the loop that we should 538 // short-circuit to. 539 540 // Split this block now, so that the loop maintains its exit block, and so 541 // that the jump from the preheader can execute the contents of the exit block 542 // without actually branching to it (the exit block should be dominated by the 543 // loop header, not the preheader). 544 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 545 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 546 547 // Okay, now we have a position to branch from and a position to branch to, 548 // insert the new conditional branch. 549 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 550 loopPreheader->getTerminator()); 551 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 552 loopPreheader->getTerminator()->eraseFromParent(); 553 554 // We need to reprocess this loop, it could be unswitched again. 555 redoLoop = true; 556 557 // Now that we know that the loop is never entered when this condition is a 558 // particular value, rewrite the loop with this info. We know that this will 559 // at least eliminate the old branch. 560 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 561 ++NumTrivial; 562 } 563 564 /// SplitExitEdges - Split all of the edges from inside the loop to their exit 565 /// blocks. Update the appropriate Phi nodes as we do so. 566 void LoopUnswitch::SplitExitEdges(Loop *L, 567 const SmallVector<BasicBlock *, 8> &ExitBlocks){ 568 569 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 570 BasicBlock *ExitBlock = ExitBlocks[i]; 571 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 572 pred_end(ExitBlock)); 573 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(), 574 ".us-lcssa", this); 575 } 576 } 577 578 /// UnswitchNontrivialCondition - We determined that the loop is profitable 579 /// to unswitch when LIC equal Val. Split it into loop versions and test the 580 /// condition outside of either loop. Return the loops created as Out1/Out2. 581 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 582 Loop *L) { 583 Function *F = loopHeader->getParent(); 584 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 585 << loopHeader->getName() << " [" << L->getBlocks().size() 586 << " blocks] in Function " << F->getName() 587 << " when '" << *Val << "' == " << *LIC << "\n"); 588 589 LoopBlocks.clear(); 590 NewBlocks.clear(); 591 592 // First step, split the preheader and exit blocks, and add these blocks to 593 // the LoopBlocks list. 594 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 595 LoopBlocks.push_back(NewPreheader); 596 597 // We want the loop to come after the preheader, but before the exit blocks. 598 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 599 600 SmallVector<BasicBlock*, 8> ExitBlocks; 601 L->getUniqueExitBlocks(ExitBlocks); 602 603 // Split all of the edges from inside the loop to their exit blocks. Update 604 // the appropriate Phi nodes as we do so. 605 SplitExitEdges(L, ExitBlocks); 606 607 // The exit blocks may have been changed due to edge splitting, recompute. 608 ExitBlocks.clear(); 609 L->getUniqueExitBlocks(ExitBlocks); 610 611 // Add exit blocks to the loop blocks. 612 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 613 614 // Next step, clone all of the basic blocks that make up the loop (including 615 // the loop preheader and exit blocks), keeping track of the mapping between 616 // the instructions and blocks. 617 NewBlocks.reserve(LoopBlocks.size()); 618 ValueMap<const Value*, Value*> VMap; 619 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 620 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); 621 NewBlocks.push_back(NewBB); 622 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. 623 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); 624 } 625 626 // Splice the newly inserted blocks into the function right before the 627 // original preheader. 628 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(), 629 NewBlocks[0], F->end()); 630 631 // Now we create the new Loop object for the versioned loop. 632 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); 633 Loop *ParentLoop = L->getParentLoop(); 634 if (ParentLoop) { 635 // Make sure to add the cloned preheader and exit blocks to the parent loop 636 // as well. 637 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 638 } 639 640 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 641 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]); 642 // The new exit block should be in the same loop as the old one. 643 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 644 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 645 646 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 647 "Exit block should have been split to have one successor!"); 648 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 649 650 // If the successor of the exit block had PHI nodes, add an entry for 651 // NewExit. 652 PHINode *PN; 653 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) { 654 PN = cast<PHINode>(I); 655 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 656 ValueMap<const Value *, Value*>::iterator It = VMap.find(V); 657 if (It != VMap.end()) V = It->second; 658 PN->addIncoming(V, NewExit); 659 } 660 } 661 662 // Rewrite the code to refer to itself. 663 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 664 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 665 E = NewBlocks[i]->end(); I != E; ++I) 666 RemapInstruction(I, VMap); 667 668 // Rewrite the original preheader to select between versions of the loop. 669 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 670 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 671 "Preheader splitting did not work correctly!"); 672 673 // Emit the new branch that selects between the two versions of this loop. 674 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 675 LPM->deleteSimpleAnalysisValue(OldBR, L); 676 OldBR->eraseFromParent(); 677 678 LoopProcessWorklist.push_back(NewLoop); 679 redoLoop = true; 680 681 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody 682 // deletes the instruction (for example by simplifying a PHI that feeds into 683 // the condition that we're unswitching on), we don't rewrite the second 684 // iteration. 685 WeakVH LICHandle(LIC); 686 687 // Now we rewrite the original code to know that the condition is true and the 688 // new code to know that the condition is false. 689 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); 690 691 // It's possible that simplifying one loop could cause the other to be 692 // changed to another value or a constant. If its a constant, don't simplify 693 // it. 694 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop && 695 LICHandle && !isa<Constant>(LICHandle)) 696 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); 697 } 698 699 /// RemoveFromWorklist - Remove all instances of I from the worklist vector 700 /// specified. 701 static void RemoveFromWorklist(Instruction *I, 702 std::vector<Instruction*> &Worklist) { 703 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 704 Worklist.end(), I); 705 while (WI != Worklist.end()) { 706 unsigned Offset = WI-Worklist.begin(); 707 Worklist.erase(WI); 708 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 709 } 710 } 711 712 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 713 /// program, replacing all uses with V and update the worklist. 714 static void ReplaceUsesOfWith(Instruction *I, Value *V, 715 std::vector<Instruction*> &Worklist, 716 Loop *L, LPPassManager *LPM) { 717 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 718 719 // Add uses to the worklist, which may be dead now. 720 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 721 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 722 Worklist.push_back(Use); 723 724 // Add users to the worklist which may be simplified now. 725 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 726 UI != E; ++UI) 727 Worklist.push_back(cast<Instruction>(*UI)); 728 LPM->deleteSimpleAnalysisValue(I, L); 729 RemoveFromWorklist(I, Worklist); 730 I->replaceAllUsesWith(V); 731 I->eraseFromParent(); 732 ++NumSimplify; 733 } 734 735 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 736 /// information, and remove any dead successors it has. 737 /// 738 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 739 std::vector<Instruction*> &Worklist, 740 Loop *L) { 741 if (pred_begin(BB) != pred_end(BB)) { 742 // This block isn't dead, since an edge to BB was just removed, see if there 743 // are any easy simplifications we can do now. 744 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 745 // If it has one pred, fold phi nodes in BB. 746 while (isa<PHINode>(BB->begin())) 747 ReplaceUsesOfWith(BB->begin(), 748 cast<PHINode>(BB->begin())->getIncomingValue(0), 749 Worklist, L, LPM); 750 751 // If this is the header of a loop and the only pred is the latch, we now 752 // have an unreachable loop. 753 if (Loop *L = LI->getLoopFor(BB)) 754 if (loopHeader == BB && L->contains(Pred)) { 755 // Remove the branch from the latch to the header block, this makes 756 // the header dead, which will make the latch dead (because the header 757 // dominates the latch). 758 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 759 Pred->getTerminator()->eraseFromParent(); 760 new UnreachableInst(BB->getContext(), Pred); 761 762 // The loop is now broken, remove it from LI. 763 RemoveLoopFromHierarchy(L); 764 765 // Reprocess the header, which now IS dead. 766 RemoveBlockIfDead(BB, Worklist, L); 767 return; 768 } 769 770 // If pred ends in a uncond branch, add uncond branch to worklist so that 771 // the two blocks will get merged. 772 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 773 if (BI->isUnconditional()) 774 Worklist.push_back(BI); 775 } 776 return; 777 } 778 779 DEBUG(dbgs() << "Nuking dead block: " << *BB); 780 781 // Remove the instructions in the basic block from the worklist. 782 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 783 RemoveFromWorklist(I, Worklist); 784 785 // Anything that uses the instructions in this basic block should have their 786 // uses replaced with undefs. 787 // If I is not void type then replaceAllUsesWith undef. 788 // This allows ValueHandlers and custom metadata to adjust itself. 789 if (!I->getType()->isVoidTy()) 790 I->replaceAllUsesWith(UndefValue::get(I->getType())); 791 } 792 793 // If this is the edge to the header block for a loop, remove the loop and 794 // promote all subloops. 795 if (Loop *BBLoop = LI->getLoopFor(BB)) { 796 if (BBLoop->getLoopLatch() == BB) 797 RemoveLoopFromHierarchy(BBLoop); 798 } 799 800 // Remove the block from the loop info, which removes it from any loops it 801 // was in. 802 LI->removeBlock(BB); 803 804 805 // Remove phi node entries in successors for this block. 806 TerminatorInst *TI = BB->getTerminator(); 807 SmallVector<BasicBlock*, 4> Succs; 808 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 809 Succs.push_back(TI->getSuccessor(i)); 810 TI->getSuccessor(i)->removePredecessor(BB); 811 } 812 813 // Unique the successors, remove anything with multiple uses. 814 array_pod_sort(Succs.begin(), Succs.end()); 815 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 816 817 // Remove the basic block, including all of the instructions contained in it. 818 LPM->deleteSimpleAnalysisValue(BB, L); 819 BB->eraseFromParent(); 820 // Remove successor blocks here that are not dead, so that we know we only 821 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 822 // then getting removed before we revisit them, which is badness. 823 // 824 for (unsigned i = 0; i != Succs.size(); ++i) 825 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 826 // One exception is loop headers. If this block was the preheader for a 827 // loop, then we DO want to visit the loop so the loop gets deleted. 828 // We know that if the successor is a loop header, that this loop had to 829 // be the preheader: the case where this was the latch block was handled 830 // above and headers can only have two predecessors. 831 if (!LI->isLoopHeader(Succs[i])) { 832 Succs.erase(Succs.begin()+i); 833 --i; 834 } 835 } 836 837 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 838 RemoveBlockIfDead(Succs[i], Worklist, L); 839 } 840 841 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has 842 /// become unwrapped, either because the backedge was deleted, or because the 843 /// edge into the header was removed. If the edge into the header from the 844 /// latch block was removed, the loop is unwrapped but subloops are still alive, 845 /// so they just reparent loops. If the loops are actually dead, they will be 846 /// removed later. 847 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 848 LPM->deleteLoopFromQueue(L); 849 RemoveLoopFromWorklist(L); 850 } 851 852 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 853 // the value specified by Val in the specified loop, or we know it does NOT have 854 // that value. Rewrite any uses of LIC or of properties correlated to it. 855 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 856 Constant *Val, 857 bool IsEqual) { 858 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 859 860 // FIXME: Support correlated properties, like: 861 // for (...) 862 // if (li1 < li2) 863 // ... 864 // if (li1 > li2) 865 // ... 866 867 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 868 // selects, switches. 869 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 870 std::vector<Instruction*> Worklist; 871 LLVMContext &Context = Val->getContext(); 872 873 874 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 875 // in the loop with the appropriate one directly. 876 if (IsEqual || (isa<ConstantInt>(Val) && 877 Val->getType()->isIntegerTy(1))) { 878 Value *Replacement; 879 if (IsEqual) 880 Replacement = Val; 881 else 882 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 883 !cast<ConstantInt>(Val)->getZExtValue()); 884 885 for (unsigned i = 0, e = Users.size(); i != e; ++i) 886 if (Instruction *U = cast<Instruction>(Users[i])) { 887 if (!L->contains(U)) 888 continue; 889 U->replaceUsesOfWith(LIC, Replacement); 890 Worklist.push_back(U); 891 } 892 SimplifyCode(Worklist, L); 893 return; 894 } 895 896 // Otherwise, we don't know the precise value of LIC, but we do know that it 897 // is certainly NOT "Val". As such, simplify any uses in the loop that we 898 // can. This case occurs when we unswitch switch statements. 899 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 900 Instruction *U = cast<Instruction>(Users[i]); 901 if (!L->contains(U)) 902 continue; 903 904 Worklist.push_back(U); 905 906 // TODO: We could do other simplifications, for example, turning 907 // 'icmp eq LIC, Val' -> false. 908 909 // If we know that LIC is not Val, use this info to simplify code. 910 SwitchInst *SI = dyn_cast<SwitchInst>(U); 911 if (SI == 0 || !isa<ConstantInt>(Val)) continue; 912 913 unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val)); 914 if (DeadCase == 0) continue; // Default case is live for multiple values. 915 916 // Found a dead case value. Don't remove PHI nodes in the 917 // successor if they become single-entry, those PHI nodes may 918 // be in the Users list. 919 920 // FIXME: This is a hack. We need to keep the successor around 921 // and hooked up so as to preserve the loop structure, because 922 // trying to update it is complicated. So instead we preserve the 923 // loop structure and put the block on a dead code path. 924 BasicBlock *Switch = SI->getParent(); 925 SplitEdge(Switch, SI->getSuccessor(DeadCase), this); 926 // Compute the successors instead of relying on the return value 927 // of SplitEdge, since it may have split the switch successor 928 // after PHI nodes. 929 BasicBlock *NewSISucc = SI->getSuccessor(DeadCase); 930 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 931 // Create an "unreachable" destination. 932 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 933 Switch->getParent(), 934 OldSISucc); 935 new UnreachableInst(Context, Abort); 936 // Force the new case destination to branch to the "unreachable" 937 // block while maintaining a (dead) CFG edge to the old block. 938 NewSISucc->getTerminator()->eraseFromParent(); 939 BranchInst::Create(Abort, OldSISucc, 940 ConstantInt::getTrue(Context), NewSISucc); 941 // Release the PHI operands for this edge. 942 for (BasicBlock::iterator II = NewSISucc->begin(); 943 PHINode *PN = dyn_cast<PHINode>(II); ++II) 944 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 945 UndefValue::get(PN->getType())); 946 // Tell the domtree about the new block. We don't fully update the 947 // domtree here -- instead we force it to do a full recomputation 948 // after the pass is complete -- but we do need to inform it of 949 // new blocks. 950 if (DT) 951 DT->addNewBlock(Abort, NewSISucc); 952 } 953 954 SimplifyCode(Worklist, L); 955 } 956 957 /// SimplifyCode - Okay, now that we have simplified some instructions in the 958 /// loop, walk over it and constant prop, dce, and fold control flow where 959 /// possible. Note that this is effectively a very simple loop-structure-aware 960 /// optimizer. During processing of this loop, L could very well be deleted, so 961 /// it must not be used. 962 /// 963 /// FIXME: When the loop optimizer is more mature, separate this out to a new 964 /// pass. 965 /// 966 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 967 while (!Worklist.empty()) { 968 Instruction *I = Worklist.back(); 969 Worklist.pop_back(); 970 971 // Simple constant folding. 972 if (Constant *C = ConstantFoldInstruction(I)) { 973 ReplaceUsesOfWith(I, C, Worklist, L, LPM); 974 continue; 975 } 976 977 // Simple DCE. 978 if (isInstructionTriviallyDead(I)) { 979 DEBUG(dbgs() << "Remove dead instruction '" << *I); 980 981 // Add uses to the worklist, which may be dead now. 982 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 983 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 984 Worklist.push_back(Use); 985 LPM->deleteSimpleAnalysisValue(I, L); 986 RemoveFromWorklist(I, Worklist); 987 I->eraseFromParent(); 988 ++NumSimplify; 989 continue; 990 } 991 992 // See if instruction simplification can hack this up. This is common for 993 // things like "select false, X, Y" after unswitching made the condition be 994 // 'false'. 995 if (Value *V = SimplifyInstruction(I)) { 996 ReplaceUsesOfWith(I, V, Worklist, L, LPM); 997 continue; 998 } 999 1000 // Special case hacks that appear commonly in unswitched code. 1001 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1002 if (BI->isUnconditional()) { 1003 // If BI's parent is the only pred of the successor, fold the two blocks 1004 // together. 1005 BasicBlock *Pred = BI->getParent(); 1006 BasicBlock *Succ = BI->getSuccessor(0); 1007 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1008 if (!SinglePred) continue; // Nothing to do. 1009 assert(SinglePred == Pred && "CFG broken"); 1010 1011 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1012 << Succ->getName() << "\n"); 1013 1014 // Resolve any single entry PHI nodes in Succ. 1015 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1016 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1017 1018 // Move all of the successor contents from Succ to Pred. 1019 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1020 Succ->end()); 1021 LPM->deleteSimpleAnalysisValue(BI, L); 1022 BI->eraseFromParent(); 1023 RemoveFromWorklist(BI, Worklist); 1024 1025 // If Succ has any successors with PHI nodes, update them to have 1026 // entries coming from Pred instead of Succ. 1027 Succ->replaceAllUsesWith(Pred); 1028 1029 // Remove Succ from the loop tree. 1030 LI->removeBlock(Succ); 1031 LPM->deleteSimpleAnalysisValue(Succ, L); 1032 Succ->eraseFromParent(); 1033 ++NumSimplify; 1034 continue; 1035 } 1036 1037 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1038 // Conditional branch. Turn it into an unconditional branch, then 1039 // remove dead blocks. 1040 continue; // FIXME: Enable. 1041 1042 DEBUG(dbgs() << "Folded branch: " << *BI); 1043 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1044 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1045 DeadSucc->removePredecessor(BI->getParent(), true); 1046 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1047 LPM->deleteSimpleAnalysisValue(BI, L); 1048 BI->eraseFromParent(); 1049 RemoveFromWorklist(BI, Worklist); 1050 ++NumSimplify; 1051 1052 RemoveBlockIfDead(DeadSucc, Worklist, L); 1053 } 1054 continue; 1055 } 1056 } 1057 } 1058