LoopUnswitch.cpp revision 02720242103ab7adc44580328bd8d35fa555bcb1
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 // We can never unswitch on vector conditions. 173 if (Cond->getType()->isVectorTy()) 174 return 0; 175 176 // Constants should be folded, not unswitched on! 177 if (isa<Constant>(Cond)) return 0; 178 179 // TODO: Handle: br (VARIANT|INVARIANT). 180 181 // Hoist simple values out. 182 if (L->makeLoopInvariant(Cond, Changed)) 183 return Cond; 184 185 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 186 if (BO->getOpcode() == Instruction::And || 187 BO->getOpcode() == Instruction::Or) { 188 // If either the left or right side is invariant, we can unswitch on this, 189 // which will cause the branch to go away in one loop and the condition to 190 // simplify in the other one. 191 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 192 return LHS; 193 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 194 return RHS; 195 } 196 197 return 0; 198} 199 200bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 201 LI = &getAnalysis<LoopInfo>(); 202 LPM = &LPM_Ref; 203 DF = getAnalysisIfAvailable<DominanceFrontier>(); 204 DT = getAnalysisIfAvailable<DominatorTree>(); 205 currentLoop = L; 206 Function *F = currentLoop->getHeader()->getParent(); 207 bool Changed = false; 208 do { 209 assert(currentLoop->isLCSSAForm(*DT)); 210 redoLoop = false; 211 Changed |= processCurrentLoop(); 212 } while(redoLoop); 213 214 if (Changed) { 215 // FIXME: Reconstruct dom info, because it is not preserved properly. 216 if (DT) 217 DT->runOnFunction(*F); 218 if (DF) 219 DF->runOnFunction(*F); 220 } 221 return Changed; 222} 223 224/// processCurrentLoop - Do actual work and unswitch loop if possible 225/// and profitable. 226bool LoopUnswitch::processCurrentLoop() { 227 bool Changed = false; 228 LLVMContext &Context = currentLoop->getHeader()->getContext(); 229 230 // Loop over all of the basic blocks in the loop. If we find an interior 231 // block that is branching on a loop-invariant condition, we can unswitch this 232 // loop. 233 for (Loop::block_iterator I = currentLoop->block_begin(), 234 E = currentLoop->block_end(); 235 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/// 292static 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. 325static 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/// 346bool 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. 408bool 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.NeverInline) { 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 ValueMap. 461// 462static inline void RemapInstruction(Instruction *I, 463 DenseMap<const Value *, Value*> &ValueMap) { 464 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 465 Value *Op = I->getOperand(op); 466 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 467 if (It != ValueMap.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. 474static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 475 LoopInfo *LI, LPPassManager *LPM) { 476 Loop *New = new Loop(); 477 478 LPM->insertLoop(New, PL); 479 480 // Add all of the blocks in L to the new loop. 481 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 482 I != E; ++I) 483 if (LI->getLoopFor(*I) == L) 484 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 485 486 // Add all of the subloops to the new loop. 487 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 488 CloneLoop(*I, New, VM, LI, LPM); 489 490 return New; 491} 492 493/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 494/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 495/// code immediately before InsertPt. 496void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 497 BasicBlock *TrueDest, 498 BasicBlock *FalseDest, 499 Instruction *InsertPt) { 500 // Insert a conditional branch on LIC to the two preheaders. The original 501 // code is the true version and the new code is the false version. 502 Value *BranchVal = LIC; 503 if (!isa<ConstantInt>(Val) || 504 Val->getType() != Type::getInt1Ty(LIC->getContext())) 505 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp"); 506 else if (Val != ConstantInt::getTrue(Val->getContext())) 507 // We want to enter the new loop when the condition is true. 508 std::swap(TrueDest, FalseDest); 509 510 // Insert the new branch. 511 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 512 513 // If either edge is critical, split it. This helps preserve LoopSimplify 514 // form for enclosing loops. 515 SplitCriticalEdge(BI, 0, this); 516 SplitCriticalEdge(BI, 1, this); 517} 518 519/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 520/// condition in it (a cond branch from its header block to its latch block, 521/// where the path through the loop that doesn't execute its body has no 522/// side-effects), unswitch it. This doesn't involve any code duplication, just 523/// moving the conditional branch outside of the loop and updating loop info. 524void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 525 Constant *Val, 526 BasicBlock *ExitBlock) { 527 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 528 << loopHeader->getName() << " [" << L->getBlocks().size() 529 << " blocks] in Function " << L->getHeader()->getParent()->getName() 530 << " on cond: " << *Val << " == " << *Cond << "\n"); 531 532 // First step, split the preheader, so that we know that there is a safe place 533 // to insert the conditional branch. We will change loopPreheader to have a 534 // conditional branch on Cond. 535 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 536 537 // Now that we have a place to insert the conditional branch, create a place 538 // to branch to: this is the exit block out of the loop that we should 539 // short-circuit to. 540 541 // Split this block now, so that the loop maintains its exit block, and so 542 // that the jump from the preheader can execute the contents of the exit block 543 // without actually branching to it (the exit block should be dominated by the 544 // loop header, not the preheader). 545 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 546 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 547 548 // Okay, now we have a position to branch from and a position to branch to, 549 // insert the new conditional branch. 550 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 551 loopPreheader->getTerminator()); 552 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 553 loopPreheader->getTerminator()->eraseFromParent(); 554 555 // We need to reprocess this loop, it could be unswitched again. 556 redoLoop = true; 557 558 // Now that we know that the loop is never entered when this condition is a 559 // particular value, rewrite the loop with this info. We know that this will 560 // at least eliminate the old branch. 561 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 562 ++NumTrivial; 563} 564 565/// SplitExitEdges - Split all of the edges from inside the loop to their exit 566/// blocks. Update the appropriate Phi nodes as we do so. 567void LoopUnswitch::SplitExitEdges(Loop *L, 568 const SmallVector<BasicBlock *, 8> &ExitBlocks) 569{ 570 571 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 572 BasicBlock *ExitBlock = ExitBlocks[i]; 573 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 574 pred_end(ExitBlock)); 575 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(), 576 ".us-lcssa", this); 577 } 578} 579 580/// UnswitchNontrivialCondition - We determined that the loop is profitable 581/// to unswitch when LIC equal Val. Split it into loop versions and test the 582/// condition outside of either loop. Return the loops created as Out1/Out2. 583void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 584 Loop *L) { 585 Function *F = loopHeader->getParent(); 586 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 587 << loopHeader->getName() << " [" << L->getBlocks().size() 588 << " blocks] in Function " << F->getName() 589 << " when '" << *Val << "' == " << *LIC << "\n"); 590 591 LoopBlocks.clear(); 592 NewBlocks.clear(); 593 594 // First step, split the preheader and exit blocks, and add these blocks to 595 // the LoopBlocks list. 596 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 597 LoopBlocks.push_back(NewPreheader); 598 599 // We want the loop to come after the preheader, but before the exit blocks. 600 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 601 602 SmallVector<BasicBlock*, 8> ExitBlocks; 603 L->getUniqueExitBlocks(ExitBlocks); 604 605 // Split all of the edges from inside the loop to their exit blocks. Update 606 // the appropriate Phi nodes as we do so. 607 SplitExitEdges(L, ExitBlocks); 608 609 // The exit blocks may have been changed due to edge splitting, recompute. 610 ExitBlocks.clear(); 611 L->getUniqueExitBlocks(ExitBlocks); 612 613 // Add exit blocks to the loop blocks. 614 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 615 616 // Next step, clone all of the basic blocks that make up the loop (including 617 // the loop preheader and exit blocks), keeping track of the mapping between 618 // the instructions and blocks. 619 NewBlocks.reserve(LoopBlocks.size()); 620 DenseMap<const Value*, Value*> ValueMap; 621 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 622 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 623 NewBlocks.push_back(New); 624 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 625 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L); 626 } 627 628 // Splice the newly inserted blocks into the function right before the 629 // original preheader. 630 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 631 NewBlocks[0], F->end()); 632 633 // Now we create the new Loop object for the versioned loop. 634 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); 635 Loop *ParentLoop = L->getParentLoop(); 636 if (ParentLoop) { 637 // Make sure to add the cloned preheader and exit blocks to the parent loop 638 // as well. 639 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 640 } 641 642 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 643 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 644 // The new exit block should be in the same loop as the old one. 645 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 646 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 647 648 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 649 "Exit block should have been split to have one successor!"); 650 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 651 652 // If the successor of the exit block had PHI nodes, add an entry for 653 // NewExit. 654 PHINode *PN; 655 for (BasicBlock::iterator I = ExitSucc->begin(); 656 (PN = dyn_cast<PHINode>(I)); ++I) { 657 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 658 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 659 if (It != ValueMap.end()) V = It->second; 660 PN->addIncoming(V, NewExit); 661 } 662 } 663 664 // Rewrite the code to refer to itself. 665 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 666 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 667 E = NewBlocks[i]->end(); I != E; ++I) 668 RemapInstruction(I, ValueMap); 669 670 // Rewrite the original preheader to select between versions of the loop. 671 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 672 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 673 "Preheader splitting did not work correctly!"); 674 675 // Emit the new branch that selects between the two versions of this loop. 676 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 677 LPM->deleteSimpleAnalysisValue(OldBR, L); 678 OldBR->eraseFromParent(); 679 680 LoopProcessWorklist.push_back(NewLoop); 681 redoLoop = true; 682 683 // Now we rewrite the original code to know that the condition is true and the 684 // new code to know that the condition is false. 685 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 686 687 // It's possible that simplifying one loop could cause the other to be 688 // deleted. If so, don't simplify it. 689 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 690 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 691 692} 693 694/// RemoveFromWorklist - Remove all instances of I from the worklist vector 695/// specified. 696static void RemoveFromWorklist(Instruction *I, 697 std::vector<Instruction*> &Worklist) { 698 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 699 Worklist.end(), I); 700 while (WI != Worklist.end()) { 701 unsigned Offset = WI-Worklist.begin(); 702 Worklist.erase(WI); 703 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 704 } 705} 706 707/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 708/// program, replacing all uses with V and update the worklist. 709static void ReplaceUsesOfWith(Instruction *I, Value *V, 710 std::vector<Instruction*> &Worklist, 711 Loop *L, LPPassManager *LPM) { 712 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 713 714 // Add uses to the worklist, which may be dead now. 715 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 716 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 717 Worklist.push_back(Use); 718 719 // Add users to the worklist which may be simplified now. 720 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 721 UI != E; ++UI) 722 Worklist.push_back(cast<Instruction>(*UI)); 723 LPM->deleteSimpleAnalysisValue(I, L); 724 RemoveFromWorklist(I, Worklist); 725 I->replaceAllUsesWith(V); 726 I->eraseFromParent(); 727 ++NumSimplify; 728} 729 730/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 731/// information, and remove any dead successors it has. 732/// 733void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 734 std::vector<Instruction*> &Worklist, 735 Loop *L) { 736 if (pred_begin(BB) != pred_end(BB)) { 737 // This block isn't dead, since an edge to BB was just removed, see if there 738 // are any easy simplifications we can do now. 739 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 740 // If it has one pred, fold phi nodes in BB. 741 while (isa<PHINode>(BB->begin())) 742 ReplaceUsesOfWith(BB->begin(), 743 cast<PHINode>(BB->begin())->getIncomingValue(0), 744 Worklist, L, LPM); 745 746 // If this is the header of a loop and the only pred is the latch, we now 747 // have an unreachable loop. 748 if (Loop *L = LI->getLoopFor(BB)) 749 if (loopHeader == BB && L->contains(Pred)) { 750 // Remove the branch from the latch to the header block, this makes 751 // the header dead, which will make the latch dead (because the header 752 // dominates the latch). 753 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 754 Pred->getTerminator()->eraseFromParent(); 755 new UnreachableInst(BB->getContext(), Pred); 756 757 // The loop is now broken, remove it from LI. 758 RemoveLoopFromHierarchy(L); 759 760 // Reprocess the header, which now IS dead. 761 RemoveBlockIfDead(BB, Worklist, L); 762 return; 763 } 764 765 // If pred ends in a uncond branch, add uncond branch to worklist so that 766 // the two blocks will get merged. 767 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 768 if (BI->isUnconditional()) 769 Worklist.push_back(BI); 770 } 771 return; 772 } 773 774 DEBUG(dbgs() << "Nuking dead block: " << *BB); 775 776 // Remove the instructions in the basic block from the worklist. 777 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 778 RemoveFromWorklist(I, Worklist); 779 780 // Anything that uses the instructions in this basic block should have their 781 // uses replaced with undefs. 782 // If I is not void type then replaceAllUsesWith undef. 783 // This allows ValueHandlers and custom metadata to adjust itself. 784 if (!I->getType()->isVoidTy()) 785 I->replaceAllUsesWith(UndefValue::get(I->getType())); 786 } 787 788 // If this is the edge to the header block for a loop, remove the loop and 789 // promote all subloops. 790 if (Loop *BBLoop = LI->getLoopFor(BB)) { 791 if (BBLoop->getLoopLatch() == BB) 792 RemoveLoopFromHierarchy(BBLoop); 793 } 794 795 // Remove the block from the loop info, which removes it from any loops it 796 // was in. 797 LI->removeBlock(BB); 798 799 800 // Remove phi node entries in successors for this block. 801 TerminatorInst *TI = BB->getTerminator(); 802 SmallVector<BasicBlock*, 4> Succs; 803 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 804 Succs.push_back(TI->getSuccessor(i)); 805 TI->getSuccessor(i)->removePredecessor(BB); 806 } 807 808 // Unique the successors, remove anything with multiple uses. 809 array_pod_sort(Succs.begin(), Succs.end()); 810 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 811 812 // Remove the basic block, including all of the instructions contained in it. 813 LPM->deleteSimpleAnalysisValue(BB, L); 814 BB->eraseFromParent(); 815 // Remove successor blocks here that are not dead, so that we know we only 816 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 817 // then getting removed before we revisit them, which is badness. 818 // 819 for (unsigned i = 0; i != Succs.size(); ++i) 820 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 821 // One exception is loop headers. If this block was the preheader for a 822 // loop, then we DO want to visit the loop so the loop gets deleted. 823 // We know that if the successor is a loop header, that this loop had to 824 // be the preheader: the case where this was the latch block was handled 825 // above and headers can only have two predecessors. 826 if (!LI->isLoopHeader(Succs[i])) { 827 Succs.erase(Succs.begin()+i); 828 --i; 829 } 830 } 831 832 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 833 RemoveBlockIfDead(Succs[i], Worklist, L); 834} 835 836/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 837/// become unwrapped, either because the backedge was deleted, or because the 838/// edge into the header was removed. If the edge into the header from the 839/// latch block was removed, the loop is unwrapped but subloops are still alive, 840/// so they just reparent loops. If the loops are actually dead, they will be 841/// removed later. 842void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 843 LPM->deleteLoopFromQueue(L); 844 RemoveLoopFromWorklist(L); 845} 846 847// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 848// the value specified by Val in the specified loop, or we know it does NOT have 849// that value. Rewrite any uses of LIC or of properties correlated to it. 850void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 851 Constant *Val, 852 bool IsEqual) { 853 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 854 855 // FIXME: Support correlated properties, like: 856 // for (...) 857 // if (li1 < li2) 858 // ... 859 // if (li1 > li2) 860 // ... 861 862 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 863 // selects, switches. 864 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 865 std::vector<Instruction*> Worklist; 866 LLVMContext &Context = Val->getContext(); 867 868 869 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 870 // in the loop with the appropriate one directly. 871 if (IsEqual || (isa<ConstantInt>(Val) && 872 Val->getType()->isIntegerTy(1))) { 873 Value *Replacement; 874 if (IsEqual) 875 Replacement = Val; 876 else 877 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 878 !cast<ConstantInt>(Val)->getZExtValue()); 879 880 for (unsigned i = 0, e = Users.size(); i != e; ++i) 881 if (Instruction *U = cast<Instruction>(Users[i])) { 882 if (!L->contains(U)) 883 continue; 884 U->replaceUsesOfWith(LIC, Replacement); 885 Worklist.push_back(U); 886 } 887 } else { 888 // Otherwise, we don't know the precise value of LIC, but we do know that it 889 // is certainly NOT "Val". As such, simplify any uses in the loop that we 890 // can. This case occurs when we unswitch switch statements. 891 for (unsigned i = 0, e = Users.size(); i != e; ++i) 892 if (Instruction *U = cast<Instruction>(Users[i])) { 893 if (!L->contains(U)) 894 continue; 895 896 Worklist.push_back(U); 897 898 // If we know that LIC is not Val, use this info to simplify code. 899 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 900 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 901 if (SI->getCaseValue(i) == Val) { 902 // Found a dead case value. Don't remove PHI nodes in the 903 // successor if they become single-entry, those PHI nodes may 904 // be in the Users list. 905 906 // FIXME: This is a hack. We need to keep the successor around 907 // and hooked up so as to preserve the loop structure, because 908 // trying to update it is complicated. So instead we preserve the 909 // loop structure and put the block on a dead code path. 910 BasicBlock *Switch = SI->getParent(); 911 SplitEdge(Switch, SI->getSuccessor(i), this); 912 // Compute the successors instead of relying on the return value 913 // of SplitEdge, since it may have split the switch successor 914 // after PHI nodes. 915 BasicBlock *NewSISucc = SI->getSuccessor(i); 916 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 917 // Create an "unreachable" destination. 918 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 919 Switch->getParent(), 920 OldSISucc); 921 new UnreachableInst(Context, Abort); 922 // Force the new case destination to branch to the "unreachable" 923 // block while maintaining a (dead) CFG edge to the old block. 924 NewSISucc->getTerminator()->eraseFromParent(); 925 BranchInst::Create(Abort, OldSISucc, 926 ConstantInt::getTrue(Context), NewSISucc); 927 // Release the PHI operands for this edge. 928 for (BasicBlock::iterator II = NewSISucc->begin(); 929 PHINode *PN = dyn_cast<PHINode>(II); ++II) 930 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 931 UndefValue::get(PN->getType())); 932 // Tell the domtree about the new block. We don't fully update the 933 // domtree here -- instead we force it to do a full recomputation 934 // after the pass is complete -- but we do need to inform it of 935 // new blocks. 936 if (DT) 937 DT->addNewBlock(Abort, NewSISucc); 938 break; 939 } 940 } 941 } 942 943 // TODO: We could do other simplifications, for example, turning 944 // LIC == Val -> false. 945 } 946 } 947 948 SimplifyCode(Worklist, L); 949} 950 951/// SimplifyCode - Okay, now that we have simplified some instructions in the 952/// loop, walk over it and constant prop, dce, and fold control flow where 953/// possible. Note that this is effectively a very simple loop-structure-aware 954/// optimizer. During processing of this loop, L could very well be deleted, so 955/// it must not be used. 956/// 957/// FIXME: When the loop optimizer is more mature, separate this out to a new 958/// pass. 959/// 960void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 961 while (!Worklist.empty()) { 962 Instruction *I = Worklist.back(); 963 Worklist.pop_back(); 964 965 // Simple constant folding. 966 if (Constant *C = ConstantFoldInstruction(I)) { 967 ReplaceUsesOfWith(I, C, Worklist, L, LPM); 968 continue; 969 } 970 971 // Simple DCE. 972 if (isInstructionTriviallyDead(I)) { 973 DEBUG(dbgs() << "Remove dead instruction '" << *I); 974 975 // Add uses to the worklist, which may be dead now. 976 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 977 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 978 Worklist.push_back(Use); 979 LPM->deleteSimpleAnalysisValue(I, L); 980 RemoveFromWorklist(I, Worklist); 981 I->eraseFromParent(); 982 ++NumSimplify; 983 continue; 984 } 985 986 // Special case hacks that appear commonly in unswitched code. 987 switch (I->getOpcode()) { 988 case Instruction::Select: 989 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 990 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L, 991 LPM); 992 continue; 993 } 994 break; 995 case Instruction::And: 996 if (isa<ConstantInt>(I->getOperand(0)) && 997 // constant -> RHS 998 I->getOperand(0)->getType()->isIntegerTy(1)) 999 cast<BinaryOperator>(I)->swapOperands(); 1000 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1001 if (CB->getType()->isIntegerTy(1)) { 1002 if (CB->isOne()) // X & 1 -> X 1003 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1004 else // X & 0 -> 0 1005 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1006 continue; 1007 } 1008 break; 1009 case Instruction::Or: 1010 if (isa<ConstantInt>(I->getOperand(0)) && 1011 // constant -> RHS 1012 I->getOperand(0)->getType()->isIntegerTy(1)) 1013 cast<BinaryOperator>(I)->swapOperands(); 1014 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1015 if (CB->getType()->isIntegerTy(1)) { 1016 if (CB->isOne()) // X | 1 -> 1 1017 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1018 else // X | 0 -> X 1019 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1020 continue; 1021 } 1022 break; 1023 case Instruction::Br: { 1024 BranchInst *BI = cast<BranchInst>(I); 1025 if (BI->isUnconditional()) { 1026 // If BI's parent is the only pred of the successor, fold the two blocks 1027 // together. 1028 BasicBlock *Pred = BI->getParent(); 1029 BasicBlock *Succ = BI->getSuccessor(0); 1030 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1031 if (!SinglePred) continue; // Nothing to do. 1032 assert(SinglePred == Pred && "CFG broken"); 1033 1034 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1035 << Succ->getName() << "\n"); 1036 1037 // Resolve any single entry PHI nodes in Succ. 1038 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1039 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1040 1041 // Move all of the successor contents from Succ to Pred. 1042 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1043 Succ->end()); 1044 LPM->deleteSimpleAnalysisValue(BI, L); 1045 BI->eraseFromParent(); 1046 RemoveFromWorklist(BI, Worklist); 1047 1048 // If Succ has any successors with PHI nodes, update them to have 1049 // entries coming from Pred instead of Succ. 1050 Succ->replaceAllUsesWith(Pred); 1051 1052 // Remove Succ from the loop tree. 1053 LI->removeBlock(Succ); 1054 LPM->deleteSimpleAnalysisValue(Succ, L); 1055 Succ->eraseFromParent(); 1056 ++NumSimplify; 1057 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1058 // Conditional branch. Turn it into an unconditional branch, then 1059 // remove dead blocks. 1060 break; // FIXME: Enable. 1061 1062 DEBUG(dbgs() << "Folded branch: " << *BI); 1063 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1064 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1065 DeadSucc->removePredecessor(BI->getParent(), true); 1066 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1067 LPM->deleteSimpleAnalysisValue(BI, L); 1068 BI->eraseFromParent(); 1069 RemoveFromWorklist(BI, Worklist); 1070 ++NumSimplify; 1071 1072 RemoveBlockIfDead(DeadSucc, Worklist, L); 1073 } 1074 break; 1075 } 1076 } 1077 } 1078} 1079