LoopUnswitch.cpp revision 11acaa374cdcebb161bf0de5f244265d78a749c1
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/LoopInfo.h" 38#include "llvm/Analysis/LoopPass.h" 39#include "llvm/Analysis/Dominators.h" 40#include "llvm/Transforms/Utils/Cloning.h" 41#include "llvm/Transforms/Utils/Local.h" 42#include "llvm/Transforms/Utils/BasicBlockUtils.h" 43#include "llvm/ADT/Statistic.h" 44#include "llvm/ADT/SmallPtrSet.h" 45#include "llvm/ADT/STLExtras.h" 46#include "llvm/Support/CommandLine.h" 47#include "llvm/Support/Debug.h" 48#include "llvm/Support/raw_ostream.h" 49#include <algorithm> 50#include <set> 51using namespace llvm; 52 53STATISTIC(NumBranches, "Number of branches unswitched"); 54STATISTIC(NumSwitches, "Number of switches unswitched"); 55STATISTIC(NumSelects , "Number of selects unswitched"); 56STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 57STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 58 59// The specific value of 50 here was chosen based only on intuition and a 60// few specific examples. 61static cl::opt<unsigned> 62Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 63 cl::init(50), cl::Hidden); 64 65namespace { 66 class LoopUnswitch : public LoopPass { 67 LoopInfo *LI; // Loop information 68 LPPassManager *LPM; 69 70 // LoopProcessWorklist - Used to check if second loop needs processing 71 // after RewriteLoopBodyWithConditionConstant rewrites first loop. 72 std::vector<Loop*> LoopProcessWorklist; 73 SmallPtrSet<Value *,8> UnswitchedVals; 74 75 bool OptimizeForSize; 76 bool redoLoop; 77 78 Loop *currentLoop; 79 DominanceFrontier *DF; 80 DominatorTree *DT; 81 BasicBlock *loopHeader; 82 BasicBlock *loopPreheader; 83 84 // LoopBlocks contains all of the basic blocks of the loop, including the 85 // preheader of the loop, the body of the loop, and the exit blocks of the 86 // loop, in that order. 87 std::vector<BasicBlock*> LoopBlocks; 88 // NewBlocks contained cloned copy of basic blocks from LoopBlocks. 89 std::vector<BasicBlock*> NewBlocks; 90 91 public: 92 static char ID; // Pass ID, replacement for typeid 93 explicit LoopUnswitch(bool Os = false) : 94 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false), 95 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL), 96 loopPreheader(NULL) {} 97 98 bool runOnLoop(Loop *L, LPPassManager &LPM); 99 bool processCurrentLoop(); 100 101 /// This transformation requires natural loop information & requires that 102 /// loop preheaders be inserted into the CFG... 103 /// 104 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 105 AU.addRequiredID(LoopSimplifyID); 106 AU.addPreservedID(LoopSimplifyID); 107 AU.addRequired<LoopInfo>(); 108 AU.addPreserved<LoopInfo>(); 109 AU.addRequiredID(LCSSAID); 110 AU.addPreservedID(LCSSAID); 111 AU.addPreserved<DominatorTree>(); 112 AU.addPreserved<DominanceFrontier>(); 113 } 114 115 private: 116 117 virtual void releaseMemory() { 118 UnswitchedVals.clear(); 119 } 120 121 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 122 /// remove it. 123 void RemoveLoopFromWorklist(Loop *L) { 124 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 125 LoopProcessWorklist.end(), L); 126 if (I != LoopProcessWorklist.end()) 127 LoopProcessWorklist.erase(I); 128 } 129 130 void initLoopData() { 131 loopHeader = currentLoop->getHeader(); 132 loopPreheader = currentLoop->getLoopPreheader(); 133 } 134 135 /// Split all of the edges from inside the loop to their exit blocks. 136 /// Update the appropriate Phi nodes as we do so. 137 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks); 138 139 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); 140 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 141 BasicBlock *ExitBlock); 142 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 143 144 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 145 Constant *Val, bool isEqual); 146 147 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 148 BasicBlock *TrueDest, 149 BasicBlock *FalseDest, 150 Instruction *InsertPt); 151 152 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 153 void RemoveBlockIfDead(BasicBlock *BB, 154 std::vector<Instruction*> &Worklist, Loop *l); 155 void RemoveLoopFromHierarchy(Loop *L); 156 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, 157 BasicBlock **LoopExit = 0); 158 159 }; 160} 161char LoopUnswitch::ID = 0; 162static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 163 164Pass *llvm::createLoopUnswitchPass(bool Os) { 165 return new LoopUnswitch(Os); 166} 167 168/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 169/// invariant in the loop, or has an invariant piece, return the invariant. 170/// Otherwise, return null. 171static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 172 // Constants should be folded, not unswitched on! 173 if (isa<Constant>(Cond)) return 0; 174 175 // TODO: Handle: br (VARIANT|INVARIANT). 176 177 // Hoist simple values out. 178 if (L->makeLoopInvariant(Cond, Changed)) 179 return Cond; 180 181 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 182 if (BO->getOpcode() == Instruction::And || 183 BO->getOpcode() == Instruction::Or) { 184 // If either the left or right side is invariant, we can unswitch on this, 185 // which will cause the branch to go away in one loop and the condition to 186 // simplify in the other one. 187 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 188 return LHS; 189 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 190 return RHS; 191 } 192 193 return 0; 194} 195 196bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 197 LI = &getAnalysis<LoopInfo>(); 198 LPM = &LPM_Ref; 199 DF = getAnalysisIfAvailable<DominanceFrontier>(); 200 DT = getAnalysisIfAvailable<DominatorTree>(); 201 currentLoop = L; 202 Function *F = currentLoop->getHeader()->getParent(); 203 bool Changed = false; 204 do { 205 assert(currentLoop->isLCSSAForm()); 206 redoLoop = false; 207 Changed |= processCurrentLoop(); 208 } while(redoLoop); 209 210 if (Changed) { 211 // FIXME: Reconstruct dom info, because it is not preserved properly. 212 if (DT) 213 DT->runOnFunction(*F); 214 if (DF) 215 DF->runOnFunction(*F); 216 } 217 return Changed; 218} 219 220/// processCurrentLoop - Do actual work and unswitch loop if possible 221/// and profitable. 222bool LoopUnswitch::processCurrentLoop() { 223 bool Changed = false; 224 LLVMContext &Context = currentLoop->getHeader()->getContext(); 225 226 // Loop over all of the basic blocks in the loop. If we find an interior 227 // block that is branching on a loop-invariant condition, we can unswitch this 228 // loop. 229 for (Loop::block_iterator I = currentLoop->block_begin(), 230 E = currentLoop->block_end(); 231 I != E; ++I) { 232 TerminatorInst *TI = (*I)->getTerminator(); 233 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 234 // If this isn't branching on an invariant condition, we can't unswitch 235 // it. 236 if (BI->isConditional()) { 237 // See if this, or some part of it, is loop invariant. If so, we can 238 // unswitch on it if we desire. 239 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 240 currentLoop, Changed); 241 if (LoopCond && UnswitchIfProfitable(LoopCond, 242 ConstantInt::getTrue(Context))) { 243 ++NumBranches; 244 return true; 245 } 246 } 247 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 248 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 249 currentLoop, Changed); 250 if (LoopCond && SI->getNumCases() > 1) { 251 // Find a value to unswitch on: 252 // FIXME: this should chose the most expensive case! 253 Constant *UnswitchVal = SI->getCaseValue(1); 254 // Do not process same value again and again. 255 if (!UnswitchedVals.insert(UnswitchVal)) 256 continue; 257 258 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 259 ++NumSwitches; 260 return true; 261 } 262 } 263 } 264 265 // Scan the instructions to check for unswitchable values. 266 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 267 BBI != E; ++BBI) 268 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 269 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 270 currentLoop, Changed); 271 if (LoopCond && UnswitchIfProfitable(LoopCond, 272 ConstantInt::getTrue(Context))) { 273 ++NumSelects; 274 return true; 275 } 276 } 277 } 278 return Changed; 279} 280 281/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 282/// 1. Exit the loop with no side effects. 283/// 2. Branch to the latch block with no side-effects. 284/// 285/// If these conditions are true, we return true and set ExitBB to the block we 286/// exit through. 287/// 288static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 289 BasicBlock *&ExitBB, 290 std::set<BasicBlock*> &Visited) { 291 if (!Visited.insert(BB).second) { 292 // Already visited and Ok, end of recursion. 293 return true; 294 } else if (!L->contains(BB)) { 295 // Otherwise, this is a loop exit, this is fine so long as this is the 296 // first exit. 297 if (ExitBB != 0) return false; 298 ExitBB = BB; 299 return true; 300 } 301 302 // Otherwise, this is an unvisited intra-loop node. Check all successors. 303 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 304 // Check to see if the successor is a trivial loop exit. 305 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 306 return false; 307 } 308 309 // Okay, everything after this looks good, check to make sure that this block 310 // doesn't include any side effects. 311 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 312 if (I->mayHaveSideEffects()) 313 return false; 314 315 return true; 316} 317 318/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 319/// leads to an exit from the specified loop, and has no side-effects in the 320/// process. If so, return the block that is exited to, otherwise return null. 321static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 322 std::set<BasicBlock*> Visited; 323 Visited.insert(L->getHeader()); // Branches to header are ok. 324 BasicBlock *ExitBB = 0; 325 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 326 return ExitBB; 327 return 0; 328} 329 330/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 331/// trivial: that is, that the condition controls whether or not the loop does 332/// anything at all. If this is a trivial condition, unswitching produces no 333/// code duplications (equivalently, it produces a simpler loop and a new empty 334/// loop, which gets deleted). 335/// 336/// If this is a trivial condition, return true, otherwise return false. When 337/// returning true, this sets Cond and Val to the condition that controls the 338/// trivial condition: when Cond dynamically equals Val, the loop is known to 339/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 340/// Cond == Val. 341/// 342bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 343 BasicBlock **LoopExit) { 344 BasicBlock *Header = currentLoop->getHeader(); 345 TerminatorInst *HeaderTerm = Header->getTerminator(); 346 LLVMContext &Context = Header->getContext(); 347 348 BasicBlock *LoopExitBB = 0; 349 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 350 // If the header block doesn't end with a conditional branch on Cond, we 351 // can't handle it. 352 if (!BI->isConditional() || BI->getCondition() != Cond) 353 return false; 354 355 // Check to see if a successor of the branch is guaranteed to go to the 356 // latch block or exit through a one exit block without having any 357 // side-effects. If so, determine the value of Cond that causes it to do 358 // this. 359 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 360 BI->getSuccessor(0)))) { 361 if (Val) *Val = ConstantInt::getTrue(Context); 362 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 363 BI->getSuccessor(1)))) { 364 if (Val) *Val = ConstantInt::getFalse(Context); 365 } 366 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 367 // If this isn't a switch on Cond, we can't handle it. 368 if (SI->getCondition() != Cond) return false; 369 370 // Check to see if a successor of the switch is guaranteed to go to the 371 // latch block or exit through a one exit block without having any 372 // side-effects. If so, determine the value of Cond that causes it to do 373 // this. Note that we can't trivially unswitch on the default case. 374 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 375 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 376 SI->getSuccessor(i)))) { 377 // Okay, we found a trivial case, remember the value that is trivial. 378 if (Val) *Val = SI->getCaseValue(i); 379 break; 380 } 381 } 382 383 // If we didn't find a single unique LoopExit block, or if the loop exit block 384 // contains phi nodes, this isn't trivial. 385 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 386 return false; // Can't handle this. 387 388 if (LoopExit) *LoopExit = LoopExitBB; 389 390 // We already know that nothing uses any scalar values defined inside of this 391 // loop. As such, we just have to check to see if this loop will execute any 392 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 393 // part of the loop that the code *would* execute. We already checked the 394 // tail, check the header now. 395 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 396 if (I->mayHaveSideEffects()) 397 return false; 398 return true; 399} 400 401/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 402/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 403/// unswitch the loop, reprocess the pieces, then return true. 404bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){ 405 406 initLoopData(); 407 408 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 409 if (!loopPreheader) 410 return false; 411 412 Function *F = loopHeader->getParent(); 413 414 // If the condition is trivial, always unswitch. There is no code growth for 415 // this case. 416 if (!IsTrivialUnswitchCondition(LoopCond)) { 417 // Check to see if it would be profitable to unswitch current loop. 418 419 // Do not do non-trivial unswitch while optimizing for size. 420 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) 421 return false; 422 423 // FIXME: This is overly conservative because it does not take into 424 // consideration code simplification opportunities and code that can 425 // be shared by the resultant unswitched loops. 426 CodeMetrics Metrics; 427 for (Loop::block_iterator I = currentLoop->block_begin(), 428 E = currentLoop->block_end(); 429 I != E; ++I) 430 Metrics.analyzeBasicBlock(*I); 431 432 // Limit the number of instructions to avoid causing significant code 433 // expansion, and the number of basic blocks, to avoid loops with 434 // large numbers of branches which cause loop unswitching to go crazy. 435 // This is a very ad-hoc heuristic. 436 if (Metrics.NumInsts > Threshold || 437 Metrics.NumBlocks * 5 > Threshold || 438 Metrics.NeverInline) { 439 DEBUG(dbgs() << "NOT unswitching loop %" 440 << currentLoop->getHeader()->getName() << ", cost too high: " 441 << currentLoop->getBlocks().size() << "\n"); 442 return false; 443 } 444 } 445 446 Constant *CondVal; 447 BasicBlock *ExitBlock; 448 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 449 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 450 } else { 451 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 452 } 453 454 return true; 455} 456 457// RemapInstruction - Convert the instruction operands from referencing the 458// current values into those specified by ValueMap. 459// 460static inline void RemapInstruction(Instruction *I, 461 DenseMap<const Value *, Value*> &ValueMap) { 462 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 463 Value *Op = I->getOperand(op); 464 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 465 if (It != ValueMap.end()) Op = It->second; 466 I->setOperand(op, Op); 467 } 468} 469 470/// CloneLoop - Recursively clone the specified loop and all of its children, 471/// mapping the blocks with the specified map. 472static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 473 LoopInfo *LI, LPPassManager *LPM) { 474 Loop *New = new Loop(); 475 476 LPM->insertLoop(New, PL); 477 478 // Add all of the blocks in L to the new loop. 479 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 480 I != E; ++I) 481 if (LI->getLoopFor(*I) == L) 482 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 483 484 // Add all of the subloops to the new loop. 485 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 486 CloneLoop(*I, New, VM, LI, LPM); 487 488 return New; 489} 490 491/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 492/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 493/// code immediately before InsertPt. 494void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 495 BasicBlock *TrueDest, 496 BasicBlock *FalseDest, 497 Instruction *InsertPt) { 498 // Insert a conditional branch on LIC to the two preheaders. The original 499 // code is the true version and the new code is the false version. 500 Value *BranchVal = LIC; 501 if (!isa<ConstantInt>(Val) || 502 Val->getType() != Type::getInt1Ty(LIC->getContext())) 503 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp"); 504 else if (Val != ConstantInt::getTrue(Val->getContext())) 505 // We want to enter the new loop when the condition is true. 506 std::swap(TrueDest, FalseDest); 507 508 // Insert the new branch. 509 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 510 511 // If either edge is critical, split it. This helps preserve LoopSimplify 512 // form for enclosing loops. 513 SplitCriticalEdge(BI, 0, this); 514 SplitCriticalEdge(BI, 1, this); 515} 516 517/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 518/// condition in it (a cond branch from its header block to its latch block, 519/// where the path through the loop that doesn't execute its body has no 520/// side-effects), unswitch it. This doesn't involve any code duplication, just 521/// moving the conditional branch outside of the loop and updating loop info. 522void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 523 Constant *Val, 524 BasicBlock *ExitBlock) { 525 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 526 << loopHeader->getName() << " [" << L->getBlocks().size() 527 << " blocks] in Function " << L->getHeader()->getParent()->getName() 528 << " on cond: " << *Val << " == " << *Cond << "\n"); 529 530 // First step, split the preheader, so that we know that there is a safe place 531 // to insert the conditional branch. We will change loopPreheader to have a 532 // conditional branch on Cond. 533 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 534 535 // Now that we have a place to insert the conditional branch, create a place 536 // to branch to: this is the exit block out of the loop that we should 537 // short-circuit to. 538 539 // Split this block now, so that the loop maintains its exit block, and so 540 // that the jump from the preheader can execute the contents of the exit block 541 // without actually branching to it (the exit block should be dominated by the 542 // loop header, not the preheader). 543 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 544 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 545 546 // Okay, now we have a position to branch from and a position to branch to, 547 // insert the new conditional branch. 548 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 549 loopPreheader->getTerminator()); 550 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 551 loopPreheader->getTerminator()->eraseFromParent(); 552 553 // We need to reprocess this loop, it could be unswitched again. 554 redoLoop = true; 555 556 // Now that we know that the loop is never entered when this condition is a 557 // particular value, rewrite the loop with this info. We know that this will 558 // at least eliminate the old branch. 559 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 560 ++NumTrivial; 561} 562 563/// SplitExitEdges - Split all of the edges from inside the loop to their exit 564/// blocks. Update the appropriate Phi nodes as we do so. 565void LoopUnswitch::SplitExitEdges(Loop *L, 566 const SmallVector<BasicBlock *, 8> &ExitBlocks) 567{ 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. 581void 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 DenseMap<const Value*, Value*> ValueMap; 619 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 620 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 621 NewBlocks.push_back(New); 622 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 623 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L); 624 } 625 626 // Splice the newly inserted blocks into the function right before the 627 // original preheader. 628 F->getBasicBlockList().splice(LoopBlocks[0], 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(), ValueMap, 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>(ValueMap[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(); 654 (PN = dyn_cast<PHINode>(I)); ++I) { 655 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 656 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 657 if (It != ValueMap.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, ValueMap); 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 // Now we rewrite the original code to know that the condition is true and the 682 // new code to know that the condition is false. 683 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 684 685 // It's possible that simplifying one loop could cause the other to be 686 // deleted. If so, don't simplify it. 687 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 688 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 689 690} 691 692/// RemoveFromWorklist - Remove all instances of I from the worklist vector 693/// specified. 694static void RemoveFromWorklist(Instruction *I, 695 std::vector<Instruction*> &Worklist) { 696 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 697 Worklist.end(), I); 698 while (WI != Worklist.end()) { 699 unsigned Offset = WI-Worklist.begin(); 700 Worklist.erase(WI); 701 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 702 } 703} 704 705/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 706/// program, replacing all uses with V and update the worklist. 707static void ReplaceUsesOfWith(Instruction *I, Value *V, 708 std::vector<Instruction*> &Worklist, 709 Loop *L, LPPassManager *LPM) { 710 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 711 712 // Add uses to the worklist, which may be dead now. 713 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 714 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 715 Worklist.push_back(Use); 716 717 // Add users to the worklist which may be simplified now. 718 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 719 UI != E; ++UI) 720 Worklist.push_back(cast<Instruction>(*UI)); 721 LPM->deleteSimpleAnalysisValue(I, L); 722 RemoveFromWorklist(I, Worklist); 723 I->replaceAllUsesWith(V); 724 I->eraseFromParent(); 725 ++NumSimplify; 726} 727 728/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 729/// information, and remove any dead successors it has. 730/// 731void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 732 std::vector<Instruction*> &Worklist, 733 Loop *L) { 734 if (pred_begin(BB) != pred_end(BB)) { 735 // This block isn't dead, since an edge to BB was just removed, see if there 736 // are any easy simplifications we can do now. 737 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 738 // If it has one pred, fold phi nodes in BB. 739 while (isa<PHINode>(BB->begin())) 740 ReplaceUsesOfWith(BB->begin(), 741 cast<PHINode>(BB->begin())->getIncomingValue(0), 742 Worklist, L, LPM); 743 744 // If this is the header of a loop and the only pred is the latch, we now 745 // have an unreachable loop. 746 if (Loop *L = LI->getLoopFor(BB)) 747 if (loopHeader == BB && L->contains(Pred)) { 748 // Remove the branch from the latch to the header block, this makes 749 // the header dead, which will make the latch dead (because the header 750 // dominates the latch). 751 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 752 Pred->getTerminator()->eraseFromParent(); 753 new UnreachableInst(BB->getContext(), Pred); 754 755 // The loop is now broken, remove it from LI. 756 RemoveLoopFromHierarchy(L); 757 758 // Reprocess the header, which now IS dead. 759 RemoveBlockIfDead(BB, Worklist, L); 760 return; 761 } 762 763 // If pred ends in a uncond branch, add uncond branch to worklist so that 764 // the two blocks will get merged. 765 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 766 if (BI->isUnconditional()) 767 Worklist.push_back(BI); 768 } 769 return; 770 } 771 772 DEBUG(dbgs() << "Nuking dead block: " << *BB); 773 774 // Remove the instructions in the basic block from the worklist. 775 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 776 RemoveFromWorklist(I, Worklist); 777 778 // Anything that uses the instructions in this basic block should have their 779 // uses replaced with undefs. 780 // If I is not void type then replaceAllUsesWith undef. 781 // This allows ValueHandlers and custom metadata to adjust itself. 782 if (!I->getType()->isVoidTy()) 783 I->replaceAllUsesWith(UndefValue::get(I->getType())); 784 } 785 786 // If this is the edge to the header block for a loop, remove the loop and 787 // promote all subloops. 788 if (Loop *BBLoop = LI->getLoopFor(BB)) { 789 if (BBLoop->getLoopLatch() == BB) 790 RemoveLoopFromHierarchy(BBLoop); 791 } 792 793 // Remove the block from the loop info, which removes it from any loops it 794 // was in. 795 LI->removeBlock(BB); 796 797 798 // Remove phi node entries in successors for this block. 799 TerminatorInst *TI = BB->getTerminator(); 800 SmallVector<BasicBlock*, 4> Succs; 801 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 802 Succs.push_back(TI->getSuccessor(i)); 803 TI->getSuccessor(i)->removePredecessor(BB); 804 } 805 806 // Unique the successors, remove anything with multiple uses. 807 array_pod_sort(Succs.begin(), Succs.end()); 808 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 809 810 // Remove the basic block, including all of the instructions contained in it. 811 LPM->deleteSimpleAnalysisValue(BB, L); 812 BB->eraseFromParent(); 813 // Remove successor blocks here that are not dead, so that we know we only 814 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 815 // then getting removed before we revisit them, which is badness. 816 // 817 for (unsigned i = 0; i != Succs.size(); ++i) 818 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 819 // One exception is loop headers. If this block was the preheader for a 820 // loop, then we DO want to visit the loop so the loop gets deleted. 821 // We know that if the successor is a loop header, that this loop had to 822 // be the preheader: the case where this was the latch block was handled 823 // above and headers can only have two predecessors. 824 if (!LI->isLoopHeader(Succs[i])) { 825 Succs.erase(Succs.begin()+i); 826 --i; 827 } 828 } 829 830 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 831 RemoveBlockIfDead(Succs[i], Worklist, L); 832} 833 834/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 835/// become unwrapped, either because the backedge was deleted, or because the 836/// edge into the header was removed. If the edge into the header from the 837/// latch block was removed, the loop is unwrapped but subloops are still alive, 838/// so they just reparent loops. If the loops are actually dead, they will be 839/// removed later. 840void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 841 LPM->deleteLoopFromQueue(L); 842 RemoveLoopFromWorklist(L); 843} 844 845// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 846// the value specified by Val in the specified loop, or we know it does NOT have 847// that value. Rewrite any uses of LIC or of properties correlated to it. 848void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 849 Constant *Val, 850 bool IsEqual) { 851 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 852 853 // FIXME: Support correlated properties, like: 854 // for (...) 855 // if (li1 < li2) 856 // ... 857 // if (li1 > li2) 858 // ... 859 860 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 861 // selects, switches. 862 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 863 std::vector<Instruction*> Worklist; 864 LLVMContext &Context = Val->getContext(); 865 866 867 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 868 // in the loop with the appropriate one directly. 869 if (IsEqual || (isa<ConstantInt>(Val) && 870 Val->getType()->isInteger(1))) { 871 Value *Replacement; 872 if (IsEqual) 873 Replacement = Val; 874 else 875 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 876 !cast<ConstantInt>(Val)->getZExtValue()); 877 878 for (unsigned i = 0, e = Users.size(); i != e; ++i) 879 if (Instruction *U = cast<Instruction>(Users[i])) { 880 if (!L->contains(U)) 881 continue; 882 U->replaceUsesOfWith(LIC, Replacement); 883 Worklist.push_back(U); 884 } 885 } else { 886 // Otherwise, we don't know the precise value of LIC, but we do know that it 887 // is certainly NOT "Val". As such, simplify any uses in the loop that we 888 // can. This case occurs when we unswitch switch statements. 889 for (unsigned i = 0, e = Users.size(); i != e; ++i) 890 if (Instruction *U = cast<Instruction>(Users[i])) { 891 if (!L->contains(U)) 892 continue; 893 894 Worklist.push_back(U); 895 896 // If we know that LIC is not Val, use this info to simplify code. 897 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 898 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 899 if (SI->getCaseValue(i) == Val) { 900 // Found a dead case value. Don't remove PHI nodes in the 901 // successor if they become single-entry, those PHI nodes may 902 // be in the Users list. 903 904 // FIXME: This is a hack. We need to keep the successor around 905 // and hooked up so as to preserve the loop structure, because 906 // trying to update it is complicated. So instead we preserve the 907 // loop structure and put the block on a dead code path. 908 BasicBlock *Switch = SI->getParent(); 909 SplitEdge(Switch, SI->getSuccessor(i), this); 910 // Compute the successors instead of relying on the return value 911 // of SplitEdge, since it may have split the switch successor 912 // after PHI nodes. 913 BasicBlock *NewSISucc = SI->getSuccessor(i); 914 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 915 // Create an "unreachable" destination. 916 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 917 Switch->getParent(), 918 OldSISucc); 919 new UnreachableInst(Context, Abort); 920 // Force the new case destination to branch to the "unreachable" 921 // block while maintaining a (dead) CFG edge to the old block. 922 NewSISucc->getTerminator()->eraseFromParent(); 923 BranchInst::Create(Abort, OldSISucc, 924 ConstantInt::getTrue(Context), NewSISucc); 925 // Release the PHI operands for this edge. 926 for (BasicBlock::iterator II = NewSISucc->begin(); 927 PHINode *PN = dyn_cast<PHINode>(II); ++II) 928 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 929 UndefValue::get(PN->getType())); 930 // Tell the domtree about the new block. We don't fully update the 931 // domtree here -- instead we force it to do a full recomputation 932 // after the pass is complete -- but we do need to inform it of 933 // new blocks. 934 if (DT) 935 DT->addNewBlock(Abort, NewSISucc); 936 break; 937 } 938 } 939 } 940 941 // TODO: We could do other simplifications, for example, turning 942 // LIC == Val -> false. 943 } 944 } 945 946 SimplifyCode(Worklist, L); 947} 948 949/// SimplifyCode - Okay, now that we have simplified some instructions in the 950/// loop, walk over it and constant prop, dce, and fold control flow where 951/// possible. Note that this is effectively a very simple loop-structure-aware 952/// optimizer. During processing of this loop, L could very well be deleted, so 953/// it must not be used. 954/// 955/// FIXME: When the loop optimizer is more mature, separate this out to a new 956/// pass. 957/// 958void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 959 while (!Worklist.empty()) { 960 Instruction *I = Worklist.back(); 961 Worklist.pop_back(); 962 963 // Simple constant folding. 964 if (Constant *C = ConstantFoldInstruction(I)) { 965 ReplaceUsesOfWith(I, C, Worklist, L, LPM); 966 continue; 967 } 968 969 // Simple DCE. 970 if (isInstructionTriviallyDead(I)) { 971 DEBUG(dbgs() << "Remove dead instruction '" << *I); 972 973 // Add uses to the worklist, which may be dead now. 974 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 975 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 976 Worklist.push_back(Use); 977 LPM->deleteSimpleAnalysisValue(I, L); 978 RemoveFromWorklist(I, Worklist); 979 I->eraseFromParent(); 980 ++NumSimplify; 981 continue; 982 } 983 984 // Special case hacks that appear commonly in unswitched code. 985 switch (I->getOpcode()) { 986 case Instruction::Select: 987 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 988 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L, 989 LPM); 990 continue; 991 } 992 break; 993 case Instruction::And: 994 if (isa<ConstantInt>(I->getOperand(0)) && 995 // constant -> RHS 996 I->getOperand(0)->getType()->isInteger(1)) 997 cast<BinaryOperator>(I)->swapOperands(); 998 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 999 if (CB->getType()->isInteger(1)) { 1000 if (CB->isOne()) // X & 1 -> X 1001 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1002 else // X & 0 -> 0 1003 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1004 continue; 1005 } 1006 break; 1007 case Instruction::Or: 1008 if (isa<ConstantInt>(I->getOperand(0)) && 1009 // constant -> RHS 1010 I->getOperand(0)->getType()->isInteger(1)) 1011 cast<BinaryOperator>(I)->swapOperands(); 1012 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1013 if (CB->getType()->isInteger(1)) { 1014 if (CB->isOne()) // X | 1 -> 1 1015 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1016 else // X | 0 -> X 1017 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1018 continue; 1019 } 1020 break; 1021 case Instruction::Br: { 1022 BranchInst *BI = cast<BranchInst>(I); 1023 if (BI->isUnconditional()) { 1024 // If BI's parent is the only pred of the successor, fold the two blocks 1025 // together. 1026 BasicBlock *Pred = BI->getParent(); 1027 BasicBlock *Succ = BI->getSuccessor(0); 1028 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1029 if (!SinglePred) continue; // Nothing to do. 1030 assert(SinglePred == Pred && "CFG broken"); 1031 1032 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1033 << Succ->getName() << "\n"); 1034 1035 // Resolve any single entry PHI nodes in Succ. 1036 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1037 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1038 1039 // Move all of the successor contents from Succ to Pred. 1040 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1041 Succ->end()); 1042 LPM->deleteSimpleAnalysisValue(BI, L); 1043 BI->eraseFromParent(); 1044 RemoveFromWorklist(BI, Worklist); 1045 1046 // If Succ has any successors with PHI nodes, update them to have 1047 // entries coming from Pred instead of Succ. 1048 Succ->replaceAllUsesWith(Pred); 1049 1050 // Remove Succ from the loop tree. 1051 LI->removeBlock(Succ); 1052 LPM->deleteSimpleAnalysisValue(Succ, L); 1053 Succ->eraseFromParent(); 1054 ++NumSimplify; 1055 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1056 // Conditional branch. Turn it into an unconditional branch, then 1057 // remove dead blocks. 1058 break; // FIXME: Enable. 1059 1060 DEBUG(dbgs() << "Folded branch: " << *BI); 1061 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1062 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1063 DeadSucc->removePredecessor(BI->getParent(), true); 1064 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1065 LPM->deleteSimpleAnalysisValue(BI, L); 1066 BI->eraseFromParent(); 1067 RemoveFromWorklist(BI, Worklist); 1068 ++NumSimplify; 1069 1070 RemoveBlockIfDead(DeadSucc, Worklist, L); 1071 } 1072 break; 1073 } 1074 } 1075 } 1076} 1077